Accessibility is crucial for inclusive app usability, yet developers often struggle to identify and fix app accessibility issues due to a lack of awareness, expertise, and inadequate tools. Current accessibility testing tools can identify accessibility issues but may not always provide guidance on how to address them. We introduce FixAlly, an automated tool designed to suggest source code fixes for accessibility issues detected by automated accessibility scanners. FixAlly employs a multi-agent LLM architecture to generate fix strategies, localize issues within the source code, and propose code…Apple Machine Learning Research
Introducing Amazon EKS support in Amazon SageMaker HyperPod
We are thrilled to introduce Amazon Elastic Kubernetes Service (Amazon EKS) support in Amazon SageMaker HyperPod, a purpose-built infrastructure engineered with resilience at its core. This capability allows for the seamless addition of SageMaker HyperPod managed compute to EKS clusters, using automated node and job resiliency features for foundation model (FM) development.
FMs are typically trained on large-scale compute clusters with hundreds or thousands of accelerators. Under such circumstances, hardware failures pose a significant challenge, because a single accelerator failure among thousands can halt the entire training process. For example, Meta Llama 3 405B pre-training over 54 days on 16K NVIDIA H100 Tensor Core GPUs experienced 419 unexpected interruptions, with 78% attributed to confirmed or suspected hardware issues, and with 58.7% of these interruptions being GPU-related problems, including NVLink failures and HBM3 memory failures.
Since its inception, SageMaker HyperPod was designed with a focus on managed resiliency features to mitigate such hardware failures, enabling FM builders such as Thomson Reuters, Perplexity AI, and Hugging Face to scale their FM training and inference on Slurm clusters. With the EKS support in HyperPod, you can now also benefit from the resiliency features on Kubernetes clusters by managing machine learning (ML) workloads using the HyperPod compute and managed Kubernetes control plane on the EKS cluster.
AI startups like Observea and Articul8, and enterprises like Thomson Reuters use this new feature set to manage their ML model development lifecycle:
“Through our use of SageMaker HyperPod, our customers and internal teams no longer have to worry about operating and configuring the Kubernetes control plane, and SageMaker HyperPod provides the network performance and optimized configurations to support complex HPC workloads. With Amazon EKS support in SageMaker HyperPod, we can reduce time we spent for undifferentiated heavy lifting in infrastructure management and reduce operational costs by over 30%.”
– Observea
“As a Kubernetes house, we are now thrilled to welcome the launch of Amazon EKS support for SageMaker HyperPod. This is a game changer for us as it integrates seamlessly with our existing training pipelines and makes it even easier for us to manage and operate our large-scale Kubernetes clusters. In addition, this also helps our end customers as we are now able to package and productize this capability into our GenAI platform, enabling our customers to run their own training and fine-tuning workloads in a more streamlined manner.”
– Articul8 AI
This post is designed for Kubernetes cluster administrators and ML scientists, providing an overview of the key features that SageMaker HyperPod introduces to facilitate large-scale model training on an EKS cluster.
The post is organized into the following three sections:
- Overview of Amazon EKS support in SageMaker HyperPod – This section provides a high-level overview of Amazon EKS support in SageMaker HyperPod, introducing three key resiliency features HyperPod compute provides on the EKS cluster. Additionally, this section explains how HyperPod provides a smooth developer experience for admins and scientists.
- HyperPod cluster setup and node resiliency features – This section provides a detailed guide on integrating HyperPod managed compute into your EKS cluster as Kubernetes worker nodes, emphasizing how its built-in resiliency features provide infrastructure stability. This section is especially beneficial for admins.
- Training job resiliency with the job auto resume functionality – In this section, we demonstrate how scientists can submit and manage their distributed training jobs using either the native Kubernetes CLI (kubectl) or optionally the new HyperPod CLI (hyperpod) with automatic job recovery enabled.
Overview of EKS support in SageMaker HyperPod
This section provides a high-level overview of Amazon EKS support in SageMaker HyperPod, introduces three key resiliency features HyperPod compute provides on the EKS cluster, and discusses how SageMaker HyperPod provides smooth user experiences for admins and scientists.
Architecture overview
Amazon EKS support in HyperPod supports a 1-to-1 mapping between an EKS cluster (serving as a Kubernetes control plane) and a HyperPod compute (attached as a group of worker nodes). You have three virtual private clouds (VPCs) in this architecture, hosting different types of resources:
- Amazon EKS VPC – An AWS managed VPC hosts the EKS control plane. This VPC doesn’t appear in the customer account. Amazon EKS creates a highly available endpoint for the managed Kubernetes API server that you use to communicate with your cluster (using tools like kubectl). The managed endpoint uses Network Load Balancer to load balance Kubernetes API servers.
- HyperPod VPC – An AWS managed VPC hosts the HyperPod compute. This VPC doesn’t appear in the customer account. The nodes connect to the EKS control plane through a cross-account elastic network interface (ENI).
- SageMaker user VPC – A user-managed VPC hosts resources such as Amazon FSx for Lustre, which is optionally associated with Amazon Simple Storage Service (Amazon S3) using an data repository association, on your account.
Cross-account ENIs also bridge communication between HyperPod compute instances and other AWS services on your account, such as Amazon Elastic Container Registry (Amazon ECR) and Amazon CloudWatch.
The following diagram illustrates the high-level architecture of Amazon EKS support in HyperPod.
HyperPod-managed resiliency features
Amazon EKS support in HyperPod provides the following three capabilities to make sure the cluster stays healthy and training jobs continue under unexpected interruptions:
- Deep health checks – This is a managed health check for stress testing GPUs and AWS Trainium instances, as well as performing Elastic Fabric Adapter (EFA) These checks can be run during the cluster creation, update, or node replacement phases, and can be enabled or disabled through HyperPod APIs.
- Automated node recovery – HyperPod performs managed, lightweight, and non-invasive checks, coupled with automated node replacement capability. The HyperPod monitoring agent continuously monitors and detects potential issues, including memory exhaustion, disk failures, GPU anomalies, kernel deadlocks, container runtime issues, and out-of-memory (OOM) crashes. Based on the underlying issue, the monitoring agent either replaces or reboots the node.
- Job auto resume – SageMaker HyperPod provides a job auto resume capability using the Kubeflow Training Operator for PyTorch to provide recovery and continuation of training jobs in the event of interruptions or failures. The extension makes sure the job waits and restarts after the node is replaced.
User experiences
In addition to the aforementioned managed resiliency features, SageMaker HyperPod provides smooth user experiences for both admins and scientists that are critical for managing a large cluster and running large-scale training jobs on them as part of the Amazon EKS integration:
- Admin experience – SageMaker HyperPod provides APIs and a console experience to create and manage node groups in the EKS cluster, along with the ability to SSH into the cluster nodes. SageMaker HyperPod also provides a mechanism to install additional dependencies on the cluster nodes using lifecycle scripts, and an API-based mechanism to provide cluster software updates and improve overall observability.
- Scientist experience – Along with enabling scientists to train FMs using Amazon EKS as the orchestrator, SageMaker HyperPod provides additional capabilities for scientists to effortlessly train models. With the HyperPod CLI, scientists can submit training jobs by providing a
.yamlfile and manage jobs (list, describe, view, cancel) without needing to usekubectl. Scientists can use open source tools like Kueue (a Kubernetes tool for job queuing) and adjacent SageMaker capabilities like managed MLflow to manage their experiments and training runs. Scientists can also access native SageMaker distributed training libraries that provide performance improvements by up to 20%. You can also enable SageMaker HyperPod compute with Amazon EKS support using third-party tools like KubeRay, which runs on the Kubernetes API. This allows you to bring your preferred job submission and management capabilities used with other Kubernetes clusters into your HyperPod environment.
HyperPod compute setup and node resiliency features
In this section, we provide a detailed guide on integrating HyperPod managed compute into your EKS cluster as Kubernetes worker nodes, and discuss how its built-in resiliency features provide infrastructure stability.
Prerequisites
You need to have the following in place prior to the HyperPod compute deployment:
- EKS cluster – You can associate HyperPod compute to an existing EKS cluster that satisfies the set of prerequisites. Alternatively, you can deploy a ready-made EKS cluster with a single AWS CloudFormation template. Refer the architecture guide for step-by-step setup instruction.
- Custom resources – Running multi-node distributed training requires various resources various components, such as device plugins, CSI drivers, and Training Operators, to be pre-deployed on the EKS cluster. You also need to deploy additional resources for the health monitoring agent and deep health check. HyperPodHelmCharts simplify the process using Helm, one of most commonly used package mangers for Kubernetes. Refer the developer guide for installation.
HyperPod compute setup
With the aforementioned resources successfully deployed, you’re now prepared to create the HyperPod compute. The cluster configuration is specified using a JSON file; the following code provides an example:
The provided configuration file contains two key highlights:
- “OnStartDeepHealthChecks”: [“InstanceStress”, “InstanceConnectivity”] – Instructs HyperPod to conduct a deep health check whenever new GPU or Trainium instances are added
- “NodeRecovery”: “Automatic” – Enables HyperPod’s automated node recovery functionality
You can create a HyperPod compute with the following aws command (you need version 2.17.47 or newer):
To verify the cluster status, you can use the following command:
This command displays the cluster details, including the cluster name, status, and creation time:
Alternatively, you can verify the cluster status through the SageMaker console. After a brief period, you can observe that the status for all nodes transitions to Running.
Node resiliency features
To gain further insight into the instances, you can use kubectl get nodes and examine the node labels. The sagemaker.amazonaws.com/node-health-status label reveals the life stage of each node. For instance, nodes with the ml.m5.2xlarge instance type are labeled as Schedulable, indicating that they have successfully passed the regular HyperPod health check. Conversely, nodes with the ml.p5.48xlarge instance type are labeled as Unschedulable, indicating that they have entered the initial deep health checks. The following code shows an example:
The deep health check logs are stored in the CloudWatch log group at /aws/sagemaker/Clusters/<cluster_name>/<cluster_id>. The log streams are logged at DeepHealthCheckResults/<log_stream_id>. When the deep health checks identify an issue, the output log provides detailed information, including the instance ID that failed the deep health checks and the specific failure reason. For example:
You can check the progress of the deep health check with the following values for the sagemaker.amazonaws.com/deep-health-check label on each node:
amazonaws.com/deep-health-check: InProgressamazonaws.com/deep-health-check: Passedamazonaws.com/deep-health-check: Failed
If a node fails the deep health checks, it will be replaced. Otherwise, it will be marked with the Schedulable label:
sagemaker.amazonaws.com/node-health-status: Schedulable
When you want to manually replace a specific node in your cluster, you can do so by manually modifying the label.
For complete list of resilience-related Kubernetes labels, please refer AWS documentation.
Even after the initial deep health checks, HyperPod periodically runs regular health checks. To view the health events detected by the HyperPod health monitoring agent, you can check the CloudWatch stream log:
- Example log group name –
/aws/sagemaker/Clusters/<cluster_name>/<cluster_id> - Example log stream name –
SagemakerHealthMonitoringAgent/<your_node_group_name>/<instance_id>
The SagemakerHealthMonitoringAgent log stream for each node contains only the detection events from the health monitoring agent. For example:
The deep health checks or the health monitor agent identify issues in a certain node, the node is labeled with sagemaker.amazonaws.com/node-health-status=UnschedulablePendingReplace:NoSchedule to avoid scheduling pods, and then the node is replaced or rebooted.
You can monitor the health status of HyperPod nodes through CloudWatch Container Insights, now with enhanced observability for Amazon EKS. Container Insights helps collect, aggregate, and summarize metrics and logs from containerized applications and microservices, providing detailed insights into performance, health, and status metrics for CPU, GPU, Trainium, EFA, and file system up to the container level. For the complete list of metrics tracked, see Amazon EKS and Kubernetes Container Insights metrics. With the Container Insights integration with SageMaker HyperPod, you can also check the individual node health status and the total number of schedulable and unschedulable nodes, as shown in the following screenshots.
You can find the Container Insights set up guide in Amazon EKS Support in Amazon SageMaker HyperPod Workshop.
Training job resiliency with the job auto resume functionality
In addition to infrastructure resiliency features, you can use the use job auto resume capability using the Kubeflow Training Operator for PyTorch to maintain the recovery and continuation of training jobs in the event of interruptions or failures. The job auto resume feature attempts to continue the job, whereas the HyperPod node auto recovery functionality works on resolving node failures (node reboot or replacement as needed) to minimize training downtime. This section demonstrates the job auto resume feature using a PyTorch FSDP example on the awsome-distributed-training repository.
To enable the job auto resume feature, you create a PyTorchJob with the fsdp.yaml manifest, which includes the following annotations and nodeSelector:
With the annotations sagemaker.amazonaws.com/enable-job-auto-resume: "true" and sagemaker.amazonaws.com/job-max-retry-count: "2", SageMaker HyperPod resumes interrupted training jobs up to two times and schedules the resumed jobs onto healthy nodes. These healthy nodes are identified by the node selector label sagemaker.amazonaws.com/node-health-status: Schedulable, ensuring that only nodes that have passed basic health checks and are available for running workloads are used for resumed jobs.
Submit the PyTorchJob using the kubectl command:
With the job auto resume feature enabled, if a job fails due to a hardware failure or any transient issues during training, SageMaker HyperPod initiates the node replacement workflow and restarts the job after the faulty nodes are replaced. You can verify the status of job auto resume by describing the PyTorchJob:
In the event of a hardware failure, the Kubeflow training job restarts as follows:
When you submit a training job with the HyperPod CLI, you can also request the job to be auto resumed in the following way:
Refer to config.yaml for full configuration. For other CLI options, refer to the documentation on Github repository.
Clean up
To delete your SageMaker HyperPod compute, use either the SageMaker console or the following AWS Command Line Interface (AWS CLI) command:
Cluster deletion can take a few minutes. You can confirm successful deletion after you see no clusters on the SageMaker console.
Conclusion
With the support for Amazon EKS in SageMaker HyperPod, customers who have standardized their FM development workflows on Kubernetes can adopt SageMaker HyperPod and manage their cluster resources using a familiar Kubernetes interface in SageMaker HyperPod. When training an FM, SageMaker HyperPod automatically monitors cluster health, and when an infrastructure fault such as a GPU failure occurs, SageMaker HyperPod automatically remediates the issue and restarts the training process from the last saved checkpoint, without any human intervention. Amazon EKS further enhances this capability by running deep health checks. Whenever a new instance is added to the SageMaker HyperPod compute, it undergoes a deep health check process to identify and replace potentially problematic instances. SageMaker HyperPod then automatically replaces or reboots nodes identified as faulty and resumes training processes in the event of unexpected interruptions, involving node replacement and job resubmission.
For an end-to-end tutorial on cluster management and FM training, visit the Amazon EKS Support in Amazon SageMaker HyperPod Workshop. For more information on infrastructure deployment and additional distributed training test cases, refer to the awsome-distributed-training repository. If you’re interested in deploying HyperPod with step-by-step commands, you can start from the aws-do-hyperpod repository.
About the authors
Keita Watanabe is a Senior GenAI Specialist Solutions Architect in the world-wide specialist organization at Amazon Web Services, where he helps develop machine learning solutions using OSS projects such as Slurm and Kubernetes. His background is in machine learning research and development. Prior to joining AWS, Keita worked in the ecommerce industry as a research scientist developing image retrieval systems for product search. Keita holds a PhD in Science from the University of Tokyo.
Alex Iankoulski is a full-stack software and infrastructure architect who likes to do deep, hands-on work. He is currently a Principal Solutions Architect in the world-wide specialist organization at AWS. In his role, he focuses on helping customers with the orchestration and scaling of ML and AI workloads on container-powered AWS services. He is also the author of the open source do framework and a Docker captain who loves applying container technologies to accelerate the pace of innovation while solving the world’s biggest challenges. During the past 10 years, Alex has worked on democratizing generative AI and ML, combating climate change, and making travel safer, healthcare better, and energy smarter.
Tomonori Shimomura is a Senior Solutions Architect on the Amazon SageMaker team, where he provides in-depth technical consultation to SageMaker customers and suggests product improvements to the product team. Before joining Amazon, he worked on the design and development of embedded software for video game consoles, and now he leverages his in-depth skills in cloud-side technology. In his free time, he enjoys playing video games, reading books, and writing software.
Arun Kumar Lokanatha is a Senior ML Solutions Architect with the Amazon SageMaker team. He specializes in large language model training workloads, helping customers build LLM workloads using SageMaker HyperPod, SageMaker training jobs, and SageMaker distributed training. Outside of work, he enjoys running, hiking, and cooking.
Manoj Ravi is a Senior Product Manager on the Amazon SageMaker team. He is passionate about building next-gen AI products and works on applications and tools to make foundation model development and deployment effortless for customers. He holds an MBA from the Haas School of Business and a master’s degree from Carnegie Mellon University. In his spare time, Manoj enjoys playing tennis and pursuing landscape photography.
A review of purpose-built accelerators for financial services
Data contains information, and information can be used to predict future behaviors, from the buying habits of customers to securities returns. Businesses are seeking a competitive advantage by being able to use the data they hold, apply it to their unique understanding of their business domain, and then generate actionable insights from it. The financial services industry (FSI) is no exception to this, and is a well-established producer and consumer of data and analytics. All industries have their own nuances and ways of doing business, and FSI is no exception—here, considerations such as regulation and zero-sum game competitive pressures loom large. This mostly non-technical post is written for FSI business leader personas such as the chief data officer, chief analytics officer, chief investment officer, head quant, head of research, and head of risk. These personas are faced with making strategic decisions on issues such as infrastructure investment, product roadmap, and competitive approach. The aim of this post is to level-set and inform in a rapidly advancing field, helping to understand competitive differentiators, and formulate an associated business strategy.
Accelerated computing is a generic term that is often used to refer to specialist hardware called purpose-built accelerators (PBAs). In financial services, nearly every type of activity, from quant research, to fraud prevention, to real-time trading, can benefit from reducing runtime. By performing a calculation more quickly, the user may be able to solve an equation more accurately, provide a better customer experience, or gain an informational edge over a competitor. These activities cover disparate fields such as basic data processing, analytics, and machine learning (ML). And finally, some activities, such as those involved with the latest advances in artificial intelligence (AI), are simply not practically possible, without hardware acceleration. ML is often associated with PBAs, so we start this post with an illustrative figure. The ML paradigm is learning followed by inference. Typically, learning is offline (not streaming real-time data, but historical data) on large volumes of data, whereas inference is online on small volumes of streaming data. Learning means identifying and capturing historical patterns from the data, and inference means mapping a current value to the historical pattern. PBAs, such as graphics processing units (GPUs), have an important role to play in both these phases. The following figure illustrates the idea of a large cluster of GPUs being used for learning, followed by a smaller number for inference. The distinct computational nature of the learning and inference phases means some hardware providers have developed independent solutions for each phase, whereas others have single solutions for both phases.
As shown in the preceding figure, the ML paradigm is learning (training) followed by inference. PBAs, such as GPUs, can be used for both these steps. In this example figure, features are extracted from raw historical data, which are then are fed into a neural network (NN). Due to model and data size, learning is distributed over multiple PBAs in an approach called parallelism. Labeled data is used to learn the model structure and weights. Unseen new streaming data is then applied to the model, and an inference (prediction) on that data is made.
This post starts by looking at the background of hardware accelerated computing, followed by reviewing the core technologies in this space. We then consider why and how accelerated computing is important for data processing. Then we review four important FSI use cases for accelerated computing. Key problem statements are identified and potential solutions given. The post finishes by summarizing the three key takeaways, and makes suggestions for actionable next steps.
Background on accelerated computing
CPUs are designed for processing small volumes of sequential data, whereas PBAs are suited for processing large volumes of parallel data. PBAs can perform some functions, such as some floating-point (FP) calculations, more efficiently than is possible by software running on CPUs. This can result in advantages such as reduced latency, increased throughput, and decreased energy consumption. The three types of PBAs are the easily reprogrammable chips such as GPUs, and two types of fixed-function acceleration; field-programmable gate arrays (FPGAs), and application-specific integrated circuits (ASICs). Fixed or semi-fixed function acceleration is practical when no updates are needed to the data processing logic. FPGAs are reprogrammable, albeit not very easily, whereas ASICs are custom designed fully fixed for a specific application, and not reprogrammable. As a general rule, the less user-friendly the speedup, the faster it is. In terms of resulting speedups, the approximate order is programming hardware, then programming against PBA APIs, then programming in an unmanaged language such as C++, then a managed language such as Python. Analysis of publications containing accelerated compute workloads by Zeta-Alpha shows a breakdown of 91.5% GPU PBAs, 4% other PBAs, 4% FPGA, and 0.5% ASICs. This post is focused on the easily reprogrammable PBAs.
The recent history of PBAs begins in 1999, when NVIDIA released its first product expressly marketed as a GPU, designed to accelerate computer graphics and image processing. By 2007, GPUs became more generalized computing devices, with applications across scientific computing and industry. In 2018, other forms of PBAs became available, and by 2020, PBAs were being widely used for parallel problems, such as training of NN. Examples of other PBAs now available include AWS Inferentia and AWS Trainium, Google TPU, and Graphcore IPU. Around this time, industry observers reported NVIDIA’s strategy pivoting from its traditional gaming and graphics focus to moving into scientific computing and data analytics.
The union of advances in hardware and ML has led us to the current day. Work by Hinton et al. in 2012 is now widely referred to as ML’s “Cambrian Explosion.” Although NN had been around since the 1960s and never really worked, Hinton noted three key changes. Firstly, they added more layers to their NN, improving their performance. Secondly, there was a massive increase in the volume of labeled data available for training. Thirdly, the presence of GPUs enabled the labeled data to be processed. Together, these elements lead to the start of a period of dramatic progress in ML, with NN being redubbed deep learning. In 2017, the landmark paper “Attention is all you need” was published, which laid out a new deep learning architecture based on the transformer. In order to train transformer models on internet-scale data, huge quantities of PBAs were needed. In November 2022, ChatGPT was released, a large language model (LLM) that used the transformer architecture, and is widely credited with starting the current generative AI boom.
Review of the technology
In this section, we review different components of the technology.
Parallel computing
Parallel computing refers to carrying out multiple processes simultaneously, and can be categorized according to the granularity at which parallelism is supported by the hardware. For example, a grid of connected instances, multiple processors within a single instance, multiple cores within a single processor, PBAs, or a combination of different approaches. Parallel computing uses these multiple processing elements simultaneously to solve a problem. This is accomplished by breaking the problem into independent parts so that each processing element can complete its part of the workload algorithm simultaneously. Parallelism is suited for workloads that are repetitive, fixed tasks, involving little conditional branching and often large amounts of data. It also means not all workloads are equally suitable for acceleration.
In parallel computing, the granularity of a task is a measure of the amount of communication overhead between the processing functional units. Granularity is typically split into the categories of fine-grained and coarse-grained. Fine-grained parallelism refers to a workload being split into a large number of small tasks, whereas coarse-grained refers to splitting into a small number of large tasks. The key difference between the two categories is the degree of communication and synchronization required between the processing units. A thread of execution is the smallest sequence of programmed instructions that can be managed independently by a scheduler, and is typically a component of a process. The multiple threads of a given process may be run concurrently by multithreading, while sharing resources such as memory. An application can achieve parallelism by using multithreading to split data and tasks into parallel subtasks and let the underlying architecture manage how the threads run, either concurrently on one core or in parallel on multiple cores. Here, each thread performs the same operation on different segments of memory so that they can operate in parallel. This, in turn, enables better system utilization and provides faster program execution.
Purpose built accelerators
Flynn’s taxonomy is a classification of computer architectures helpful in understanding PBAs. Two classifications of relevance are single instruction stream, multiple data streams (SIMD), and the SIMD sub-classification of single instruction, multiple thread (SIMT). SIMD describes computers with multiple processing elements that perform the same operation on multiple data points simultaneously. SIMT describes processors that are able to operate on data vectors and arrays (as opposed to just scalars), and therefore handle big data workloads efficiently. Each SIMT core has multiple threads that run in parallel, thereby giving true simultaneous parallel hardware-level execution. CPUs have a relatively small number of complex cores and are designed to run a sequence of operations (threads) as fast as possible, and can run a few tens of these threads in parallel. GPUs, in contrast, feature smaller cores and are designed to run thousands of threads in parallel in the SIMT paradigm. It is this design that primarily distinguishes GPUs from CPUs and allows GPUs to excel at regular, dense, numerical, data-flow-dominated workloads.
Suppliers of data center GPUs include NVIDIA, AMD, Intel, and others. The AWS P5 EC2 instance type range is based on the NVIDIA H100 chip, which uses the Hopper architecture. The Hopper H100 GPU (SXM5 variant) architecture includes 8 GPU processing clusters (GPCs), 66 texture processing clusters (TPCs), 2 Streaming Multiprocessors (SMs)/TPC, 528 Tensor cores/GPU, and 128 CUDA cores/SM. Additionally, it features 80 GB HBM3 GPU memory, 900 GBps NVLink GPU-to-GPU interconnect, and a 50 MB L2 cache minimizing HBM3 trips. An NVIDIA GPU is assembled in a hierarchal manner: the GPU contains multiple GPCs, and the role of each GPC is to act as a container to hold all the components together. Each GPC has a raster engine for graphics and several TPCs. Inside each TPC is a texture unit, some logic control, and multiple SMs. Inside each SM are multiple CUDA and Tensor cores, and it is here that the compute work happens. The ratio of units GPU:GPC:TPC:SM:CUDA core/Tensor core varies according to release and version. This hierarchal architecture is illustrated in the following figure.
SMs are the fundamental building blocks of an NVIDIA GPU, and consist of CUDA cores, Tensor cores, distributed shared memory, and instructions to support dynamic programming. When a CUDA program is invoked, work is distributed to the multithreaded SMs with available execution capacity. The CUDA core, released in 2007, is a GPU core approximately equal to a CPU core. Although it’s not as powerful as a CPU core, the CUDA core advantage is its ability to be used for large-scale parallel computing. Like a CPU core, each CUDA core still only runs one operation per clock cycle; however, the GPU SIMD architecture enables large numbers of CUDA cores to simultaneously address one data point each. CUDA cores are split into support for different precision, meaning that in the same clock cycle, multiple precision work can be done. The CUDA core is well suited for high-performance computing (HPC) use cases, but is not so well suited for the matrix math found in ML. The Tensor core, released in 2017, is another NVIDIA proprietary GPU core that enables mixed-precision computing, and is designed to support the matrix math of ML. Tensor cores support mixed FP accuracy matrix math in a computationally efficient manner by treating matrices as primitives and being able to perform multiple operations in one clock cycle. This makes GPUs well suited for data-heavy, matrix math-based, ML training workloads, and real-time inference workloads needing synchronicity at scale. Both use cases require the ability to move data around the chip quickly and controllably.
From 2010 onwards, other PBAs have started becoming available to consumers, such as AWS Trainium, Google’s TPU, and Graphcore’s IPU. While an in-depth review on other PBAs is beyond the scope of this post, the core principle is one of designing a chip from the ground up, based around ML-style workloads. Specifically, ML workloads are typified by irregular and sparse data access patterns. This means there is a requirement to support fine-grained parallelism based on irregular computation with aperiodic memory access patterns. Other PBAs tackle this problem statement in a variety of different ways from NVIDIA GPUs, including having cores and supporting architecture complex enough for running completely distinct programs, and decoupling thread data access from the instruction flow by having distributed memory next to the cores.
AWS accelerator hardware
AWS currently offers a range of 68 Amazon Elastic Compute Cloud (Amazon EC2) instance types for accelerated compute. Examples include F1 Xilinx FPGAs, P5 NVIDIA Hopper H100 GPUs, G4ad AMD Radeon Pro V520 GPUs, DL2q Qualcomm AI 100, DL1 Habana Gaudi, Inf2 powered by Inferentia2, and Trn1 powered by Trainium. In March 2024, AWS announced it will offer the new NVIDIA Blackwell platform, featuring the new GB200 Grace Blackwell chip. Each EC2 instance type has a number of variables associated with it, such as price, chip maker, Regional availability, amount of memory, amount of storage, and network bandwidth.
AWS chips are produced by our own Annapurna Labs team, a chip and software designer, which is a wholly owned subsidiary of Amazon. The Inferentia chip became generally available (GA) in December 2019, followed by Trainium GA in October 2022, and Inferentia2 GA in April 2023. In November 2023, AWS announced the next generation Trainium2 chip. By owning the supply and manufacturing chain, AWS is able to offer high-levels of availability of its own chips. Availability AWS Regions are shown in the following table, with more Regions coming soon. Both Inferentia2 and Trainium use the same basic components, but with differing layouts, accounting for the different workloads they are designed to support. Both chips use two NeuronCore-v2 cores each, connected by a variable number of NeuronLink-v2 interconnects. The NeuronCores contain four engines: the first three include a ScalarEngine for scalar calculations, a VectorEngine for vector calculations, and a TensorEngine for matrix calculations. By analogy to an NVIDIA GPU, the first two are comparable to CUDA cores, and the latter is equivalent to TensorCores. And finally, there is a C++ programmable GPSIMD-engine allowing for custom operations. The silicon architecture of the two chips is very similar, meaning that the same software can be used for both, minimizing changes on the user side, and this similarity can be mapped back to their two roles. In general, the learning phase of ML is typically bounded by bandwidth associated with moving large volumes of data to the chip and about the chip. The inference phase of ML is typically bounded by memory, not compute. To maximize absolute-performance and price-performance, Trainium chips have twice as many NeuronLink-v2 interconnects as Inferentia2, and Trainium instances also contain more chips per instance than Inferentia2 instances. All these differences are implemented at the server level. AWS customers such as Databricks and Anthropic use these chips to train and run their ML models.
The following figures illustrate the chip-level schematic for the architectures of Inferentia2 and Trainium.
The following table shows the metadata of three of the largest accelerated compute instances.
| Instance Name | GPU Nvidia H100 Chips | Trainium Chips | Inferentia Chips | vCPU Cores | Chip Memory (GiB) | Host Memory (GiB) | Instance Storage (TB) | Instance Bandwidth (Gbps) | EBS Bandwidth (Gbps) | PBA Chip Peer-to-Peer Bandwidth (GBps) |
| p5.48xlarge | 8 | 0 | 0 | 192 | 640 | 2048 | 8 x 3.84 SSD | 3,200 | 80 | 900 NVSwitch |
| inf2.48xlarge | 0 | 0 | 12 | 192 | 384 | 768 | EBS only | 100 | 60 | 192 NeuronLink-v2 |
| trn1n.32xlarge | 0 | 16 | 0 | 128 | 512 | 512 | 4 x 1.9 SSD | 1,600 | 80 | 768 NeuronLink-v2 |
The following table summarizes performance and cost.
| Instance Name | On-Demand Rate ($/hr) | 3Yr RI Rate ($/hr) | FP8 TFLOPS | FP16 TFLOPS | FP32 TFLOPS | $/TFLOPS (FP16, theoretical) | Source Reference |
| p5.48xlarge | 98.32 | 43.18 | 16,000 | 8,000 | 8,000 | $5.40 | URL |
| inf2.48xlarge | 12.98 | 5.19 | 2,280 | 2,280 | 570 | $2.28 | URL |
| trn1n.32xlarge | 24.78 | 9.29 | 3,040 | 3,040 | 760 | $3.06 | URL |
The following table summarizes Region availability.
| Instance Name | Number of AWS Regions Supported In | AWS Regions Supported In | Default Quota Limit |
| p5.48xlarge | 4 | us-east-2; us-east-1; us-west-2; eu-north-1 | 0 |
| inf2.48xlarge | 13 | us-east-2; us-east-1; us-west-2; ap-south-1; ap-southeast-1; ap-southeast-2; ap-northeast-1; eu-central-1; eu-west-1; eu-west-2; eu-west-3; eu-north-1; sa-east-1; | 0 |
| trn1n.32xlarge | 3 | us-east-2; us-east-1; us-west-2; eu-north-1; ap-northeast-1; ap-south-1; ap-southeast-4 | 0 |
After a user has selected the EC2 instance type, it can then be combined with AWS services designed to support large-scale accelerated computing use cases, including high-bandwidth networking (Elastic Fabric Adapter), virtualization (AWS Nitro Enclaves), hyper-scale clustering (Amazon EC2 UltraClusters), low-latency storage (Amazon FSx for Lustre), and encryption (AWS Key Management Service), while noting not all services are available for all instances in all Regions.
The following figure shows an example of a large-scale deployment of P5 EC2 instances, includes UltraCluster support for 20,000 H100 GPUs, with non-blocking petabit-scale networking, and high-throughput low latency storage. Using the same architecture, UltraCluster supports Trainium scaling to over 60,000 chips.
In summary, we see two general trends in the hardware acceleration space. Firstly, improving price-performance to handle increasing data processing volumes and model sizes, coupled with a need to serve more users, more quickly, and at reduced cost. Secondly, improving security of the associated workloads by preventing unauthorized users from being able to access training data, code, or model weights.
Accelerator software
CPUs and GPUs are designed for different types of workloads. However, CPU workloads can run on GPUs, a process called general-purpose computing on graphics processing units (GPGPU). In order to run a CPU workload on a GPU, the work needs to be reformulated in terms of graphics primitives supported by the GPU. This reformulation can be carried out manually, though it is difficult programming, requiring writing code in a low-level language to map data to graphics, process it, and then map it back. Instead, it is commonly carried out by a GPGPU software framework, allowing the programmer to ignore the underlying graphical concepts, and enabling straightforward coding against the GPU using standard programming languages such as Python. Such frameworks are designed for sequential parallelism against GPUs (or other PBAs) without requiring concurrency or threads. Examples of GPGPU frameworks are the vendor-neutral open source OpenCL and the proprietary NVIDIA CUDA.
For the Amazon PBA chips Inferentia2 and Trainium, the SDK is AWS Neuron. This SDK enables development, profiling, and deployment of workloads onto these PBAs. Neuron has various native integrations to third-party ML frameworks like PyTorch, TensorFlow, and JAX. Additionally, Neuron includes a compiler, runtime driver, as well as debug and profiling utilities. This toolset includes Neuron-top for real-time visualization of the NeuronCore and vCPU utilization, host and device memory usage, and a breakdown of memory allocation. This information is also available in JSON format if neuron-monitor is used, including Neuron-ls for device discovery and topology information. With Neuron, users can use inf2 and trn1n instances with a range of AWS compute services, such as Amazon SageMaker, Amazon Elastic Container Service, Amazon Elastic Kubernetes Service, AWS Batch, and AWS ParallelCluster. This usability, tooling, and integrations of the Neuron SDK has made Amazon PBAs extremely popular with users. For example, over 90% of the top 100 Hugging Face models (now over 100,000 AI models) now run on AWS using Optimum Neuron, enabling the Hugging Face transformer natively supported for Neuron. In summary, the Neuron SDK allows developers to easily parallelize ML algorithms, such as those commonly found in FSI. The following figure illustrates the Neuron software stack.
The CUDA API and SDK were first released by NVIDIA in 2007. CUDA offers high-level parallel programming concepts that can be compiled to the GPU, giving direct access to the GPU’s virtual instruction set and therefore the ability to specify thread-level parallelism. To achieve this, CUDA added one extension to the C language to let users declare functions that could run and compile on the GPU, and a lightweight way to call those functions. The core idea behind CUDA was to remove programmers’ barrier to entry for coding against GPUs by allowing use of existing skills and tools as much as possible, while being more user friendly than OpenCL. The CUDA platform includes drivers, runtime kernels, compilers, libraries, and developer tools. This includes a wide and impressive range of ML libraries like cuDNN and NCCL. The CUDA platform is used through complier directives and extensions to standard languages, such as the Python cuNumeric library. CUDA has continuously optimized over the years, using its proprietary nature to improve performance on NVIDIA hardware relative to vendor-neutral solutions like OpenCL. Over time, the CUDA programming paradigm and stack has become deeply embedded in all aspects of the ML ecosystem, from academia to open source ML repositories.
To date, alternative GPU platforms to CUDA have not seen widespread adoption. There are three key reasons for this. Firstly, CUDA has had a decades-long head start, and benefits from the networking effect of its mature ecosystem, from organizational inertia of change, and from risk aversion to change. Secondly, migrating CUDA code to a different GPU platform can be technically difficult, given the complexity of the ML models typically being accelerated. Thirdly, CUDA has integrations with major third-party ML libraries, such as TensorFlow and PyTorch.
Despite the central role CUDA plays in the AI/ML community, there is movement by users to diversify their accelerated workflows by movement towards a Pythonic programming layer to make training more open. A number of such efforts are underway, including projects like Triton and OneAPI, and cloud service features such as Amazon SageMaker Neo. Triton is an open source project lead by OpenAI that enables developers to use different acceleration hardware using entirely open source code. Triton uses an intermediate compiler to convert models written in supported frameworks into an intermediate representation that can then be lowered into highly optimized code for PBAs. Triton is therefore a hardware-agnostic convergence layer that hides chip differences.
Soon to be released is the AWS neuron kernel interface (NKI) programming interface. NKI is a Python-based programming environment designed for the compiler, which adopts commonly used Triton-like syntax and tile-level semantics. NKI provides customization capabilities to fully optimize performance by enabling users to write custom kernels, by passing almost all of the AWS compiler layers.
OneAPI is an open source project lead by Intel for a unified API across different accelerators including GPUs, other PBAs, and FPGAs. Intel believes that future competition in this space will happen for inference, unlike in the learning phase, where there is no software dependency. To this end, OneAPI toolkits support CUDA code migration, analysis, and debug tools. Other efforts are building on top of OneAPI; for, example the Unified Acceleration Foundation’s (UXL) goal is a new open standard accelerator software ecosystem. UXL consortium members include Intel, Google, and ARM.
Amazon SageMaker is an AWS service providing an ML development environment, where the user can select chip type from the service’s fleet of Intel, AMD, NVIDIA, and AWS hardware, offering varied cost-performance-accuracy trade-offs. Amazon contributes to Apache TVM, an open source ML compiler framework for GPUs and PBAs, enabling computations on any hardware backend. SageMaker Neo uses Apache TVM to perform static optimizations on trained models for inference for any given hardware target. Looking to the future, the accelerator software field is likely to evolve; however, this may be slow to happen.
Accelerator supply-demand imbalances
It has been widely reported for the last few years that GPUs are in short supply. Such shortages have led to industry leaders speaking out. For example, Sam Altman said “We’re so short on GPUs the less people use our products the better… we don’t have enough GPUs,” and Elon Musk said “It seems like everyone and their dog is buying GPUs at this point.”
The factors leading to this have been high demand coupled with low supply. High demand has risen from a range of sectors, including crypto mining, gaming, generic data processing, and AI. Omdia Research estimates 49% of GPUs go to the hyper-clouds (such as AWS or Azure), 27% go to big tech (such as Meta and Tesla), 20% go to GPU clouds (such as Coreweave and Lambda) and 6% go to other companies (such as OpenAI and FSI firms). The State of AI Report gives the size and owners of the largest A100 clusters, the top few being Meta with 21,400, Tesla with 16,000, XTX with 10,000, and Stability AI with 5,408. GPU supply has been limited by factors including lack of manufacturing competition and ability at all levels in the supply chain, and restricted supply of base components such as rare metals and circuit boards. Additionally, rate of manufacturing is slow, with an H100 taking 6 months to make. Socio-political events have also caused delays and issues, such as a COVID backlog, and with inert gases for manufacturing coming from Russia. A final issue impacting supply is that chip makers strategically allocate their supply to meet their long-term business objectives, which may not always align with end-users’ needs.
Supported workloads
In order to benefit from hardware acceleration, a workload needs to be parallelizable. An entire branch of science is dedicated to parallelizable problems. In The Landscape of Parallel Computing Research, 13 fields (termed dwarfs) are found to be fundamentally parallelizable, including dense and sparse linear algebra, Monte Carlo methods, and graphical models. The authors also call out a series of fields they term “embarrassingly sequential” for which the opposite holds. In FSI, one of the main data structures dealt with is time series, a series of sequential observations. Many time series algorithms have the property where each subsequent observation is dependent on previous observations. This means only some time series workloads can be efficiently computed in parallel. For example, a moving average is a good example of a computation that seems inherently sequential, but for which there is an efficient parallel algorithm. Sequential models, such as Recurrent Neural Networks (RNN) and Neural Ordinary Differential Equations, also have parallel implementations. In FSI, non-time series workloads are also underpinned by algorithms that can be parallelized. For example, Markovitz portfolio optimization requires the computationally intensive inversion of large covariance matrices, for which GPU implementations exist.
In computer science, a number can be represented with different levels of precision, such as double precision (FP64), single precision (FP32), and half-precision (FP16). Different chips support different representations, and different representations are suitable for different use cases. The lower the precision, the less storage is required, and the faster the number is to process for a given amount of computational power. FP64 is used in HPC fields, such as the natural sciences and financial modeling, resulting in minimal rounding errors. FP32 provides a balance between accuracy and speed, is used in applications such as graphics, and is the standard for GPUs. FP16 is used in deep learning where computational speed is valued, and the lower precision won’t drastically affect the model’s performance. More recently, other number representations have been developed which aim to improve the balance between acceleration and precision, such as OCP Standard FP8, Google BFloat16, and Posits. An example of a mixed representation use case is the updating of model parameters by gradient decent, part of the backpropagation algorithm, as used in deep learning. Typically this is done using FP32 to reduce rounding errors, however, in order to reduce memory load, the parameters and gradients can be stored in FP16, meaning there is a conversion requirement. In this case, BFloat16 is a good choice because it prevents float overflow errors while keeping enough precision for the algorithm to work.
As lower-precision workloads become more important, hardware and infrastructure trends are changing accordingly. For example, comparing the latest NVIDIA GB200 chip against the previous generation NVIDIA H100 chip, lower representation FP8 performance has increased 505%, but FP64 performance has only increased 265%. Likewise, in the forthcoming Trainium2 chip, the focus has been on lower-bit performance increases, giving a 400% performance increase over the previous generation. Looking to the future, we might expect to see a convergence between HPC and AI workloads, as AI starts to become increasingly important in solving what were traditionally HPC FP64 precision problems.
Accelerator benchmarking
When considering compute services, users benchmark measures such as price-performance, absolute performance, availability, latency, and throughput. Price-performance means how much compute can be done for $1, or what is the equivalent dollar cost for a given number of FP operations. For a perfect system, the price-performance ratio increases linearly as the size of a job scales up. A complicating factor when benchmarking compute grids on AWS is that EC2 instances come in a range of system parameters and a grid might contain more than one instance type, therefore systems are benchmarked at the grid level rather than on a more granular basis. Users often want to complete a job as quickly as possible and at the lowest cost; the constituent details of the system that achieves this aren’t as important.
A second benchmarking measure is absolute-performance, meaning how quickly can a given job be completed independent of price. Given linear scaling, job completion time can be reduced by simply adding more compute. However, it might be that the job isn’t infinitely divisible, and that only a single computational unit is required. In this case, the absolute performance of that computational unit is important. In an earlier section, we provided a table with one performance measure, the $/TFLOP ratio based on the chip specifications. However, as a rule of thumb, when such theoretical values are compared against experimental values, only around 45% is realized.
There are a few different ways to calculate price-performance. The first is to use a standard benchmark, such as LINPACK, HPL-MxP, or MFU (Model FLOPS Utilization). These can run a wide range of calculations that are representative of varying use cases, such as general use, HPC, and mixed HPC and AI workloads. From this, the TFLOP/s at a given FP precision for the system can be measured, along with the dollar-cost of running the system. However, it might be that the user has specific use cases in mind. In this case, the best data will come from price-performance data on a more representative benchmark.
There are various types of representative benchmark commonly seen. Firstly, the user can use real production data and applications with the hardware being benchmarked. This option gives the most reliable results, but can be difficult to achieve due to operational and compliance hurdles. Secondly, the user can replicate their existing use case with a synthetic data generator, avoiding the challenges of getting production data into new test systems. Thirdly, the use can employ a third-party benchmark for the use case, if one exists. For example, STAC is a company that coordinates an FSI community called the STAC Benchmark Council, which maintain a selection of accelerator benchmarks, including A2, A3, ML and AI (LLM). A2 is designed for compute-intensive analytic workloads involved in pricing and risk management. Specifically, the A2 workload uses option price discovery by Monte Carlo estimation of Heston-based Greeks for a path-dependent, multi-asset option with early exercise. STAC members can access A2 benchmarking reports, for example EC2 c5.metal, with the oneAPI. STAC-ML benchmarks the latency of NN inference—the time from receiving new input data until the model output is computed. STAC-A3 benchmarks the backtesting of trading algorithms to determine how strategies would have performed on historical data. This benchmark supports accelerator parallelism to run many backtesting experiments simultaneously, for the same security. For each benchmark, there exists a series of software packages (termed STAC Packs), which are accelerator-API specific. For some of the preceding benchmarks, STAC Packs are maintained by providers such as NVIDIA (CUDA) and Intel (oneAPI).
Some FSI market participants are performing in-house benchmarking at the microarchitecture level, in order to optimize performance as far as possible. Citadel has published microbenchmarks for NVIDIA GPU chips, dissecting the microarchitecture to achieve “bare-metal performance tuning,” noting that peak performance is inaccessible to software written in plain CUDA. Jane Street has looked at performance optimization through functional programming techniques, while PDT Partners has supported work on the Nixpkgs repository of ML packages using CUDA.
Some AWS customers have benchmarked the AWS PBAs against other EC2 instance types. ByteDance, the technology company that runs the video-sharing app TikTok, benchmarked Inf1 against a comparable EC2 GPU instance type. With Inf1, they were able to reduce their inference latency by 25%, and costs by 65%. In a second example, Inf2 is benchmarked against a comparable inference-optimized EC2 instance. The benchmark used is the RoBERTa-Base, a popular model used in natural language processing (NLP) applications, that uses the transformer architecture. In the following figure, on the x-axis we plotted throughput (the number of inferences that are completed in a set period of time), and on the y-axis we plotted latency (the time it takes the deep learning model to provide an output). The figure shows that Inf2 gives higher throughput and lower latency than the comparable EC2 instance type.
In a third benchmark example, Hugging Face benchmarked the trn1.32xlarge instance (16 Trainium chips) and two comparable EC2 instance types. For the first instance type, they ran fine-tuning for the BERT Large model on the full Yelp review dataset, using the BF16 data format with the maximum sequence length supported by the model (512). The benchmark results show the Trainium job is five times faster while being only 30% more expensive, resulting in a “huge improvement in cost-performance.” For the latter instance type, they ran three tests: language pretraining with GPT2, token classification with BERT Large, and image classification with the Vision Transformer. These results showed trn1 to be 2–5 times faster and 3–8 times cheaper than the comparable EC2 instance types.
FSI use cases
As with other industry sectors, there are two reasons why FSI uses acceleration. The first is to get a fixed result in the lowest time possible, for example parsing a dataset. The second is to get the best result in a fixed time, for example overnight parameter re-estimation. Use cases for acceleration exist across the FSI, including banking, capital markets, insurance, and payments. However, the most pressing demand comes from capital markets, because acceleration speeds up workloads and time is one of the easiest edges people can get in the financial markets. Put differently, a time advantage in financial services often equates to an informational advantage.
We begin by providing some definitions:
- Parsing is the process of converting between data formats
- Analytics is data processing using either deterministic or simple statistical methods
- ML is the science of learning models from data, using a variety of different methods, and then making decisions and predictions
- AI is an application able to solve problems using ML
In this section, we review some of the FSI use cases of PBAs. As many FSI activities can be parallelized, most of what is done in FSI can be sped up with PBAs. This includes most modeling, simulations, and optimization problems— currently in FSI, deep learning is only a small part of the landscape. We identify four classes of FSI use cases and look at applications in each class: parsing financial data, analytics on financial data, ML on financial data, and low-latency applications. To try and show how these classes relate to each other, the following figure shows a simplified representation of a typical capital market’s workflow. In this figure, acceleration categories have been assigned to the workflow steps. However, in reality, every step in the process may be able to benefit from one or more of the defined acceleration categories.
Parsing
A typical capital markets workflow consists of receiving data and then parsing it into a useable form. This data is commonly market data, as output from a trading venue’s matching engine, or onward from a market data vendor. Market participants who are receiving either live or historical data feeds need to ingest this data and perform one or more steps, such as parse the message out of a binary protocol, rebuild the limit order book (LOB), or combine multiple feeds into a single normalized format. Any of these parsing steps that run in parallel could be sped up relative to sequential processing. To give an idea of scale, the largest financial data feed is the consolidated US equity options feed, termed OPRA. This feed comes from 18 different trading venues, with 1.5 million contracts broadcast across 96 channels, with a supported peak message rate of 400 billion messages per day, equating to approximately 12 TB per day, or 3 PB per year. As well as maintaining real-time feeds, participants need to maintain a historical depositary, sometimes of several years in size. Processing of historical repositories is done offline, but is often a source of major cost. Overall, a large consumer of market data, such as an investment bank, might consume 200 feeds from across public and private trading venues, vendors, and redistributors.
Any point in this data processing pipeline that can be parallelized, can potentially be sped up by acceleration. For example:
- Trading venues broadcast on channels, which can be groupings of alphabetical tickers or products.
- On a given channel, different tickers update messages are broadcast sequentially. These can then be parsed out into unique streams per ticker.
- For a given LOB, some events might be applicable to individual price levels independently.
- Historical data is normally (but not always) independent inter-day, meaning that days can be parsed independently.
In GPU Accelerated Data Preparation for Limit Order Book Modeling, the authors describe a GPU pipeline handling data collection, LOB pre-processing, data normalization, and batching into training samples. The authors note their LOB pre-processing relies on the previous LOB state, and must be done sequentially. For LOB building, FPGAs seem to be used more commonly than GPUs because of the fixed nature of the workload; see examples from Xilinx and Algo-Logic. For example code for a build lab, using the AWS FPGA F1 instance type, refer to the following GitHub repo.
An important part of the data pipeline is the production of features, both online and offline. Features (also called alphas, signals, or predictors) are statistical representations of the data, which can then be used in downstream model building. A current trend in the FSI prediction space is the large-scale automation of dataset ingestion, curation, processing, feature extraction, feature combination, and model building. An example of this approach is given by WorldQuant, an algorithmic trading firm. The WSJ reports “a data group scours the globe for interesting and new data sets, including everything from detailed market pricing data to shipping statistics to footfall in stores captured by apps on smartphones”. WorldQuant states “in 2007 we had two data sets—today [2022] we have more than 1,400.” The general idea being if they could buy, consume, create, and web scrape more data than anyone else, they could create more alphas, and find more opportunities. Such an approach is based on performance being proportional to √N, where N is the number of alphas. Therefore, as long as an alpha is not perfectly correlated with another, there is value in adding it to the set. In 2010, WorldQuant was producing several thousand alphas per year, by 2016 had one million alphas, by 2022, had multiple millions, with a stated ambition to get to 100 million alphas. Although traditional quant finance mandates the importance of an economic rationale behind an alpha, the data-driven approach is led purely by the patterns in the data. After alphas have been produced, they can be intelligently merged together in a time-variant manner. Examples of signal combination methodologies which can benefit from PBA speed-up include Mean Variance Optimization and Bayesian Model Averaging. The same WSJ article states “No one alpha is important. Our edge is putting things together, it’s the implementation…. The idea is that with so many ‘alphas,’ even weak signals can be useful. If counting cars in parking lots next to big box retailers has only a tiny predictive power for those retailers’ stock prices, it can still be used to enhance a bigger prediction if combined with other weak signals. For example, an uptick in cars at Walmart parking lots—itself a relatively weak signal—could combine with similar trends captured by mobile phone apps and credit-card receipts harvested by companies that scan emails to create a more reliable prediction.” The automated process of data ingestion, processing, packaging, combination, and prediction is referred to by WorldQuant as their “alpha factory.”
From examples such as those we’ve discussed, it seems clear that parallelization, speed-up and scale-up, of such huge data pipelines is potentially an important differentiator. All the way through this pipeline, activities could be accelerated using PBAs. For example, for use at the signal combination phase, the Shapley value is a metric that can be used to compute the contribution of a given feature to a prediction. Shapley value computation has PBA-acceleration support in the Python XGBoost library.
Analytics
In this section, we consider the applicability of accelerator parallelism to analytics workloads. One of the parallelizable dwarfs is Monte Carlo, and for FSI and time series work in general, this is an important method. Monte Carlo is a way to compute expected values by generating random scenarios and then averaging them. By using GPUs, a simulated path can be assigned to each thread, allowing simulation of thousands of paths in parallel.
Post the 2008 credit crunch, new regulations require banks to run credit valuation adjustment (CVA) calculations every 24 hours. CVA is an adjustment to a derivatives price as charged by a bank to a counterparty. CVA is one of a family of related valuation adjustments collectively known as xVA, which include debt valuation adjustment (DVA), initial margin valuation adjustment (MVA), capital valuation adjustment (KVA), and funding valuation adjustment (FVA). Because this adjustment calculation can happen over large portfolios of complex, non-linear instruments, closed-form analytical solutions aren’t possible, and as such an empirical approximation by a technique such as Monte Carlo is required. The downside of Monte Carlo here is how computationally demanding it is, due to the size of the search space. The advent of this new regulation coincided with the coming of age of GPUs, and as such banks commonly use GPU grids to run their xVA calculations. In XVA principles, nested Monte Carlo strategies, and GPU optimizations, the authors find a nested simulation time of about an hour for a billion scenarios on the bank portfolio, and a GPU speedup of 100 times faster relative to CPUs. Rather than develop xVA applications internally, banks often use third-party independent software vendor (ISV) solutions to run their xVA calculations, such as Murex M3 or S&P Global XVA. Banking customers can choose to run such ISV software as a service (SaaS) solutions inside their own AWS accounts, and often on AWS accelerated instances.
A second use of PBAs in FSI Monte Carlo is in option pricing, especially for exotic options whose payoff is sometimes too complex to solve in closed-form. The core idea is using a random number generator (RNG) to simulate the stochastic components in a formula and then average the results, leading to the expected value. The more paths that are simulated, the more accurate the result is. In Quasi-Monte Carlo methods for calculating derivatives sensitivities on the GPU, the authors find 200-times greater speedup over CPUs, and additionally develop a number of refinements to reduce variance, leading to fewer paths needing to be simulated. In High Performance Financial Simulation Using Randomized Quasi-Monte Carlo Methods, the authors survey quasi Monte Carlo sequences in GPU libraries and review commercial software tools to help migrate Monte Carlo pricing models to GPU. In GPU Computing in Bayesian Inference of Realized Stochastic Volatility Model, the author computes a volatility measure using Hybrid Monte Carlo (HMC) applied to realized stochastic volatility (RSV), parallelized on a GPU, resulting in a 17-times faster speedup. Finally, in Derivatives Sensitivities Computation under Heston Model on GPU, the authors achieve a 200-times faster speedup; however, the accuracy of the GPU method is inferior for some Greeks relative to CPU.
A third use of PBAs in FSI Monte Carlo is in LOB simulations. We can categorize different types of LOB simulations: replay of the public historical data, replay of the mapped public-private historical data, replay of synthetic LOB data, and replay of a mix of historical and synthetic data to simulate the effects of a feedback loop. For each of these types of simulation, there are multiple ways in which hardware acceleration could occur. For example, for the simple replay case, each accelerator thread could have a different LOB. For the synthetic data case, each thread could have a different version of the same LOB, thereby allowing multiple realizations of a single LOB. In Limit Order Book Simulations: A Review, the authors provide their own simulator classification scheme based on the mathematical modeling technique used—point processes, agent based, deep learning, stochastic differential equations. In JAX-LOB: A GPU-Accelerated limit order book simulator to unlock large scale reinforcement learning for trading, the authors use GPU accelerated training, processing thousands of LOBs in parallel, giving a “notably reduced per message processing time.”
Machine learning
Generative AI is the most topical ML application at this point in time. Generative AI has four main applications: classification, prediction, understanding, and data generation, which in turn map to use cases such as customer experience, knowledge worker productivity, surfacing information and sentiment, and innovation and automation. FSI examples exist for all of these; however, a thorough review of these is beyond the scope of this post. For this post, we remain focused on PBA applicability and look at two of these topics: chatbots and time series prediction.
The 2017, the publication of the paper Attention is all you need resulted in a new wave of interest in ML. The transformer architecture presented in this paper allowed for a highly parallelizable network structure, meaning more data could be processed than before, allowing patterns to be better captured. This has driven impressive real-world performance, as seen by popular public foundation models (FMs) such as OpenAI ChatGPT, and Anthropic Claude. These factors in turn have driven new demand for PBAs for training and inference on these models.
FMs, also termed LLMs, or chatbots when text focused, are models that are typically trained on a broad spectrum of generalized and unlabeled data and are capable of performing a wide variety of general tasks in FSI, such as the Bridgewater Associates LLM-powered Investment Analyst Assistant, which generates charts, computes financial indicators, and summarizes results. FSI LLMs are reviewed in Large Language Models in Finance: A Survey and A Survey of Large Language Models for Financial Applications: Progress, Prospects and Challenges. FMs are often used as base models for developing more specialized downstream applications.
PBAs are used in three different types of FM training. Firstly, to train a FM from scratch. In BloombergGPT: A Large Language Model for Finance, the training dataset was 51% financial data from their systems and 49% public data, such as Wikipedia and Pile. SageMaker was used to train and evaluate their FM. Specifically, 64 p4d.24xlarge instances, giving a total of 512 A100 GPUs. Also used was SageMaker model parallelism, enabling the automatic distribution of the large model across multiple GPU devices and instances. The authors started with a compute budget of 1.3 million GPU hours, and noted training took approximately 53 days.
The second training approach is to fine-tune an existing FM. This requires using an FM whose model parameters are exposed, and updating them in light of new data. This approach can be effective when the data corpus differs significantly from the FM training data. Fine-tuning is cheaper and quicker than training FM from scratch, because the volume of data is likely to be much smaller. As with the larger-scale training from scratch, fine-tuning benefits significantly from hardware acceleration. In an FSI example, Efficient Continual Pre-training for Building Domain Specific Large Language Models, the authors fine-tune an FM and find that their approach outperforms standard continual pre-training performance with just 10% of the corpus size and cost, without any degradation on open-domain standard tasks.
The third training approach is to perform Retrieval Augmented Generation (RAG). To equip FMs with up-to-date and proprietary information, organizations use RAG, a technique that fetches data from company data sources and enriches the prompt to provide more relevant and accurate responses. The two-step workflow consists of ingesting data and vectorizing data, followed by runtime orchestration. Although hardware acceleration is less common in RAG applications, latency of search is a key component and as such the inference step of RAG can be hardware optimized. For example, the performance of OpenSearch, a vectorized database available on AWS, can be improved by using PBAs, with both NVIDIA GPUs and Inferentia being supported.
For these three training approaches, the role of PBAs varies. For processing the huge data volumes of FM building, PBAs are essential. Then, as the training volumes reduce, so does the value-add role of the PBA. Independent of how the model has been trained, PBAs have a key role in LLM inference, again because they are optimized for memory bandwidth and parallelism. The specifics of how to optimally use an accelerator depend on the use case—for example, a paid-for-service chatbot might be latency sensitive, whereas for a free version, a delay of a few milliseconds might be acceptable. If a delay is acceptable, then batching the queries together could help make sure a given chip’s processes are saturated, giving better dollar usage of the resource. Dollar costs are particularly importance in inference, because unlike training, which is a one-time cost, inference is a reoccurring cost.
Using ML for financial time series prediction is nothing new; a large body of public research exists on these methods and applications dating to the 1970s and beyond—for approximately the last decade, PBAs have been applied to this field. As discussed earlier, most ML approaches can be accelerated with hardware; however, the attention-based architecture using the transformer model is currently the most topical. We consider three areas of FSI application: time series FMs, NN for securities prediction, and reinforcement learning (RL).
The initial work on LLMs was conducted on text-based models. This was followed by multi-modal models, able to handle images and other data structures. Subsequent to this, publications have started to appear on time series FMs, including Amazon Chronos, Nixtla TimeGEN-1, and Google TimesFM. The behavior of the time series models appears to be similar to that of the language models. For example, in Scaling-laws for Large Time-series Models, the authors observe the models follow the same scaling laws. A review of these models is provided in Foundation Models for Time Series Analysis: A Tutorial and Survey. As with leading LLMs, time series FMs are likely to be successfully trained on large clusters of PBAs. In terms of size, GPT-3 was trained on a cluster of 10,000 V100s. The size of the GPT-4 training cluster is not public, but is speculated to have been trained on a cluster of 10,000–25,000 A100s. This is analogous in size to one algorithmic trading firm’s statement, “our dedicated research cluster contains … 25,000 A/V100 GPUs (and growing fast).”
Looking to the future, one possible outcome might be that time series FMs, trained at huge expense by a few large corporates, become the base models for all financial prediction. Financial services firms then modify these FMs through additional training with private data or their own insights. Examples of private labeled data might be knowledge of which orders and executions in the public feed belonged to them, or similarly which (meta)orders and executions had parent-child relationships.
Although such financial time series FMs trained on PBA clusters may offer enhanced predictive capabilities, they also bring risks. For example, the EU’s AI act, adopted in March 2024, states that if a model has been trained with a total compute power in excess of 1025 FLOPs, then that model is considered to pose “systemic risk” and is subject to enhanced regulation, including fines of 3% of global turnover, so on this basis Meta announced in June 2024 they will not be enabling some models inside Europe. This legislation assumes that training compute is a direct proxy for model capabilities. EpochAI provides an analysis of the training compute required for a wide range of FMs; for example, GPT-4 took 2.125 FLOPS to train (exceeding the threshold by a factor of 2.1), whereas BloombergGPT took 2.423 FLOPS (under the threshold by a factor of 0.02). It seems possible that in the future, similar legislation may apply to financial FMs, or even to the PBA clusters themselves, with some market participants choosing not to operate in legislative regimes that are subject to such risks.
Feature engineering plays a key role in building NN models, because features are fed into the NN model. As seen earlier in this post, some participants have generated large numbers of features. Examples of features derived from market time series data include bid-ask spreads, weighted mid-points, imbalance measures, decompositions, liquidity predictions, trends, change-points, and mean-reversions. Together, the features are called the feature space. A transformer assigns more importance to part of the input feature space, even though it might only be a small part of the data. Learning which part of the data is more important than another depends on the context of the features. The true power of FMs in time series prediction is the ability to capture these conditional probabilities (the context) across the feature space. To give a simple example, based on historical data, trends might reduce in strength as they go on, leading to a change-point, and then reversion to the mean. A transformer potentially offers the ability to recognize this pattern and capture the relationship between the features more accurately than other approaches. An informative visualization of this for the textual case is given by the FT article Generative AI exists because of the transformer. In order to build and train such FMs on PBAs, access to high-quality historical data tightly coupled with scalable compute to generate the features is an essential prerequisite.
Prior to the advent of the transformer, NN have historically been applied to securities prediction with varying degrees of success. Deep Learning for Limit Order Books uses a cluster of 50 GPUs to predict the sign of the future return by mapping the price levels of the LOB to the visible input layer of a NN, resulting in a trinomial output layer. Conditional on the return the sign, the magnitude of the return is estimated using regression. Deep Learning Financial Market Data uses raw LOB data pre-processed into discrete, fixed-length features for training a recurrent autoencoder, whose recurrent structure allows learning patterns on different time scales. Inference occurs by generating the decoded LOB, and nearest-matching that to the real-time data.
In Multi-Horizon Forecasting for Limit Order Books: Novel Deep Learning Approaches and Hardware Acceleration using Intelligent Processing Units, the authors benchmark the performance of Graphcore IPUs against an NVIDIA GPU on an encoder-decoder NN model. Given that encoder-decoder models rely on recurrent neural layers, they generally suffer from slow training processes. The authors address this by finding that the IPU offers a significant training speedup over the GPU, 694% on average, analogous to the speedup a transformer architecture would provide. In some examples of post-transformer work in this space, Generative AI for End-to-End Limit Order Book Modelling and A Generative Model Of A Limit Order Book Using Recurrent Neural Networks have trained LLM analogues on historical LOB data, interpreting each LOB event (such as insertions, cancellations, and executions) as a word and predicting the series of events following a given word history. However, the authors find the prediction horizon for LOB dynamics appears to be limited to a few tens of events, possibly because of the high-dimensionality of the problem and the presence of long-range correlations in order sign. These results have been improved in the work “Microstructure Modes” — Disentangling the Joint Dynamics of Prices & Order Flow, by down-sampling the data and reducing its dimensionality, allowing identification of stable components.
RL is an ML technique where an algorithm interacts with a dynamic environment that provides feedback to the algorithm, allowing the algorithm to iteratively optimize a reward metric. Because RL closely mimics how human traders interact with the world, there are various areas of applicability in FSI. In JAX-LOB: A GPU-Accelerated limit order book simulator to unlock large scale reinforcement learning for trading, the authors use GPUs for end-to-end RL training. RL agent training with a GPU has a 7-times speedup relative to a CPU based simulation implementation. The authors then apply this to the problem of optimal trade execution. A second FSI application of RL to optimal trade execution has been reported by JPMorgan in an algorithm called LOXM.
Latency-sensitive, real-time workloads
Being able to transmit, process, and act on data more quickly than others gives an informational advantage. In the financial markets, this is directly equivalent to being able to profit from trading. These real-time, latency-sensitive workloads exist on a spectrum, from the most sensitive to the least sensitive. The specific numbers in the following table are open to debate, but present the general idea.
| Band | Latency | Application Examples |
| 1 | Less than 1 microsecond | Low-latency trading strategy. Tick 2 trade. |
| 2 | 1–4 microseconds | Feed handler. Raw or normalized format. |
| 3 | 40 microseconds | Normalized format and symbology. |
| 4 | 4–200 milliseconds | Consolidated feed. Full tick. |
| 5 | 1 second to daily | Intraday and EOD. Reference, Corp, FI, derivatives. |
The most latency-sensitive use cases are typically handled by FPGA or custom ASICs. These react to incoming network traffic, like market data, and put triggering logic directly into the network interface controller. Easily reprogrammable PBAs play little to no role in any latency sensitive work, due to the SIMD architecture being designed for the use case of parallel processing large amounts of data with a bandwidth bottleneck of getting data onto the chip.
However, three factors maybe driving change in the role hardware acceleration plays in the low-latency space. Firstly, as PBAs mature, some of their previous barriers are being reduced. For example, NVIDIA’s new NVLink design now enables significantly higher bandwidth relative to previous chip interconnects, meaning that data can get onto the chip far more quickly than before. Comparing the latest NVIDIA GB200 chip against the previous generation NVIDIA H100 chip, NVLink performance has increased 400%, from 900 GBps to 3.6 TBps.
Secondly, some observers believe the race for speed is shifting to a “race for intelligence.” With approximately only ten major firms competing in the top-tier low latency space, the barrier to entry seems almost unsurmountable for other parties. At some point, low-latency hardware and techniques might slowly diffuse through technology supplier offerings, eventually leveling the playing field, perhaps having been driven by new regulations.
Thirdly, although FPGA/ASIC undoubtedly provides the fastest performance, they come at a cost of being a drain on resources. Their developers are hard to hire for, the work has long deployment cycles, and it results in a significant maintenance burden with bugs that are difficult to diagnose and triage. Firms are keen to identify alternatives.
Although the most latency-sensitive work will remain on FPGA/ASIC, there may be a shift of less latency-sensitive work from FPGA/ASIC to GPUs and other PBAs as users weigh the trade-off between speed and other factors. In comparison, easily reprogrammable PBA processors are now simple to hire for, are straightforward to code against and maintain, and allow for relatively rapid innovation. Looking to the future, we may see innovation at the language level, for example, through functional programming with array-languages such as the Co-dfns project, as well as further innovation at the hardware level, with future chips tightly integrating the best components of today’s FPGAs, GPUs and CPUs.
Key Takeaways
In this section, we present three key takeaways. Firstly, the global supply-demand ratio for GPUs is low, meaning price can be high, but availability can be low. This can be a constraining factor for end-user businesses wanting to innovate in this space. AWS helps address this on behalf of its customers in three ways:
- Through economies of scale, AWS is able to offer significant availability of the PBAs, including GPUs.
- Through in-house research and development, AWS is able to offer its own PBAs, developed and manufactured in-house, which are not subject to the constraints of the wider market, while also having optimized price-performance.
- AWS innovates at the software level to improve allocation to the end-user. Therefore, although total capacity might be fixed, by using intelligent allocation algorithms, AWS is better able to meet customers’ needs. For example, Amazon EC2 Capacity Blocks for ML enables guaranteed access to the required PBAs at the point in time they are needed.
The second takeaway is that proprietary software can lock users in to a single supplier and end up acting as a barrier to innovation. In the case of PBAs, the chips that use proprietary software mean that users can’t easily move between chip manufacturers, as opposed to open source software supporting multiple chip manufacturers. Any future supply constraints, such as regional armed conflict, could further exasperate existing supply-demand imbalances. Although migrating existing legacy workloads from an acceleration chip with proprietary software can be challenging, new greenfield workloads can be built on open source libraries without difficulty. In the FSI space, examples of legacy workloads might include risk calculations, and examples of greenfield workloads might include time series prediction using FMs. In the long term, business leaders need to consider and formulate their strategy for moving away from software lock-in, and enable access to wider acceleration hardware offerings, with the cost benefits that can bring.
The final takeaway is that financial services, and the subsection of capital markets in particular, is subject to constant and evolving competitive pressures. Over time, the industry has seen the race for differentiation move from data access rights, to latency, and now to an increased focus on predictive power. Looking to the future, if the world of financial prediction is based in part on a small number of expensive and complex FMs built and trained by a few large global corporates, where will the differentiation come from? Speculative areas could range from at-scale feature engineering to being able to better handle increased regulatory burdens. Whichever field it comes from, it is certain to include data processing and analytics at its core, and therefore benefit from hardware acceleration.
Conclusion
This post aimed to provide business leaders with a non-technical overview of PBAs and their role within the FSI. With this technology currently being regularly discussed in the mainstream media, it is essential business leaders understand the basis of this technology and its potential future role. Nearly every organization is now looking to a data-centric future, enabled by cloud-based infrastructure and real-time analytics, to support revenue-generating AI and ML use cases. One of the ways organizations will be differentiated in this race will be by making the right strategic decisions about technologies, partners, and approaches. This includes topics such as open source versus closed source, build versus buy, tool complexity and associated ease of use, hiring and retention challenges, and price-performance. Such topics are not just technology decisions within a business, but also cultural and strategic ones.
Business leaders are encouraged to reach out to their AWS point of contact and ask how AWS can help their business win in the long term using PBAs. This might result in a range of outcomes, from a short proof of concept against an existing well-defined business problem, to a written strategy document that can be consumed and debated by peers, to onsite technical workshops and business briefing days. Whatever the outcome, the future of this space is sure to be exciting!
Acknowledgements
I would like to thank the following parties for their kind input and guidance in writing this post: Andrea Rodolico, Alex Kimber, and Shruti Koparkar. Any errors are mine alone.
About the Author
Dr. Hugh Christensen works at Amazon Web Services with a specialization in data analytics. He holds undergraduate and master’s degrees from Oxford University, the latter in computational biophysics, and a PhD in Bayesian inference from Cambridge University. Hugh’s areas of interest include time series data, data strategy, data leadership, and using analytics to drive revenue generation. You can connect with Hugh on LinkedIn.
Anomaly detection in streaming time series data with online learning using Amazon Managed Service for Apache Flink
Time series data is a distinct category that incorporates time as a fundamental element in its structure. In a time series, data points are collected sequentially, often at regular intervals, and they typically exhibit certain patterns, such as trends, seasonal variations, or cyclical behaviors. Common examples of time series data include sales revenue, system performance data (such as CPU utilization and memory usage), credit card transactions, sensor readings, and user activity analytics.
Time series anomaly detection is the process of identifying unexpected or unusual patterns in data that unfold over time. An anomaly, also known as an outlier, occurs when a data point deviates significantly from an expected pattern.
For some time series, like those with well-defined expected ranges such as machine operating temperatures or CPU usage, a threshold-based approach might suffice. However, in areas like fraud detection and sales, where simple rules fall short due to their inability to catch anomalies across complex relationships, more sophisticated techniques are required to identify unexpected occurrences.
In this post, we demonstrate how to build a robust real-time anomaly detection solution for streaming time series data using Amazon Managed Service for Apache Flink and other AWS managed services.
Solution overview
The following diagram illustrates the core architecture of the Anomaly Detection Stack solution.
This solution employs machine learning (ML) for anomaly detection, and doesn’t require users to have prior AI expertise. It offers an AWS CloudFormation template for straightforward deployment in an AWS account. With the CloudFormation template, you can deploy an application stack with the necessary AWS resources required for detecting anomalies. Setting up one stack creates an application with one anomaly detection task or detector. You can set up multiple such stacks to run them simultaneously, with each one analyzing the data and reporting back the anomalies.
The application, once deployed, constructs an ML model using the Random Cut Forest (RCF) algorithm. It initially sources input time series data from Amazon Managed Streaming for Apache Kafka (Amazon MSK) using this live stream for model training. Post-training, the model continues to process incoming data points from the stream. It evaluates these points against the historical trends of the corresponding time series. The model also generates an initial raw anomaly score while processing and maintains an internal threshold to eliminate noisy data points. Subsequently, the model generates a normalized anomaly score for each data point that the model treats as an anomaly. These scores, ranging from 0–100, indicate the deviation from typical patterns; scores closer to 100 signify higher anomaly levels. You have the flexibility to set a custom threshold on these anomaly scores, allowing you to define what you consider anomalous.
This solution uses a CloudFormation template, which takes inputs such as MSK broker endpoint and topics, AWS Identity and Access Management (IAM) roles, and other parameters related to virtual private cloud (VPC) configuration. The template creates the essential resources like the Apache Flink application and Amazon SageMaker real-time endpoint in the customer account.
To request the access to this solution, send an email to anomalydetection-support-canvas@amazon.com.
In this post, we outline how you can build an end-to-end solution with the Anomaly Detection Stack. Consider a hypothetical sales scenario where AnyBooks, an on-campus bookstore at a large university, sells various supplies to college students. Due to the timing of class schedules, their seasonality is such that they sell around 20 Item-A units and 30 Item-B units during even hours, and approximately half that during odd hours throughout the day. Recently, there have been some unexplained spikes in the quantity of items sold, and the management team wants to start tracking these quantity anomalies so that they can better plan their staffing and inventory levels.
The following diagram shows the detailed architecture for the end-to-end solution.
In the following sections, we discuss each layer shown in the preceding diagram.
Ingestion
In the ingestion layer, an AWS Lambda function retrieves sales transactions for the current minute from a PostgreSQL transactional database, transforms each record into a JSON message, and publishes it to an input Kafka topic. This Lambda function is configured to run every minute using Amazon EventBridge Scheduler.
Anomaly detection stack
The Flink application initiates the process of reading raw data from the input MSK topic, training the model, and commencing the detection of anomalies, ultimately recording them to the MSK output topic. The following code is the output results JSON:
The following is a brief explanation of the output fields:
- measure – This represents the metric we are tracking for anomalies. In our case, the
measurefield is thequantityof sales forItem-A. - aggregatedMeasureVaue – This represents the aggregated value of
quantityin the time window. - timeseriesid – This unique identifier corresponds to a combination of unique values for the dimensions and the metric. In this scenario, it’s the product name,
Item-A, within theproduct_name - anomalyConfidenceScore – As the model evolves through learning and inference, this confidence score will progressively improve.
- anomalyScore – This field represents the score for anomaly detection. With an
anomalyThresholdset at 70, any value exceeding 70 is considered a potential anomaly. - modelStage – When the model is in the learning phase, the
anomalyScoreis 0.0 and the value of this field is set toLEARNING. After the learning is complete, the value of this field changes toINFERENCE. - anomalyDecisionThreshold – The decision threshold is provided as input in the CloudFormation stack. If you determine there are too many false positives, you can increase this threshold to change the sensitivity.
- anomalyDecision – If the
anomalyScoreexceeds theanomalyDecisionThreshold, this field is set to 1, indicating an anomaly is detected.
Transform
In the transformation layer, an Amazon Data Firehose stream is configured to consume data from the output Kafka topic and invoke a Lambda function for transformation. The Lambda function flattens the nested JSON data from the Kafka topic. The transformed results are then partitioned by date and stored in an Amazon Simple Storage Service (Amazon S3) bucket in Parquet format. An AWS Glue crawler is used to crawl the data in the Amazon S3 location and catalog it in the AWS Glue Data Catalog, making it ready for querying and analysis.
Visualize
To visualize the data, we’ve created an Amazon QuickSight dashboard that connects to the data in Amazon S3 through the Data Catalog and queries it using Amazon Athena. The dashboard can be refreshed to display the latest detected anomalies, as shown in the following screenshot.
In this example, the darker blue line in the line graph represents the seasonality of the quantity measure for Item-A over time, showing higher values during even hours and lower values during odd hours. The pink line represents the anomaly detection score, plotted on the right Y-axis. The anomaly score approaches 100 when the quantity value significantly deviates from its seasonal pattern. The blue line represents the anomaly threshold, set at 70. When anomalyScore exceeds this threshold, anomalyDecision is set to 1.
The “Number of Timeseries Tracked” KPI displays how many time series the model is currently monitoring. In this case, because we’re tracking two products (Item-A and Item-B), the count is 2. The “Number of Datapoints Processed” KPI shows the total number of data points the model has processed, and the “Anomaly Confidence Score” indicates the confidence level in predicting anomalies. Initially, this score is low, but will approach 100 as the model matures over time.
Notification
Although visualization is valuable for investigating anomalies, data analysts often prefer to receive near real-time notifications for critical anomalies. This is achieved by adding a Lambda function that reads results from the output Kafka topic and analyzes them. If the anomalyScore value exceeds the defined threshold, the function invokes an Amazon Simple Notification Service (Amazon SNS) topic to send email or SMS notifications to a designated list, alerting the team about the anomaly in near real time.
Conclusion
This post demonstrated how to build a robust real-time anomaly detection solution for streaming time series data using Managed Service for Apache Flink and other AWS services. We walked through an end-to-end architecture that ingests data from a source database, passes it through an Apache Flink application that trains an ML model and detects anomalies, and then lands the anomaly data in an S3 data lake. The anomaly scores and decisions are visualized through a QuickSight dashboard connected to the Amazon S3 data using AWS Glue and Athena. Additionally, a Lambda function analyzes the results and sends notifications in near real time.
With AWS managed services like Amazon MSK, Data Firehose, Lambda, and SageMaker, you can rapidly deploy and scale this anomaly detection solution for your own time series use cases. This allows you to automatically identify unexpected behaviors or patterns in your data streams in real time without manual rules or thresholds.
Give this solution a try, and explore how real-time anomaly detection on AWS can unlock insights and optimize operations across your business!
About the Authors
Noah Soprala is a Solutions Architect based out of Dallas. He is a trusted advisor to his customers and helps them build innovative solutions using AWS technologies. Noah has over 20 years of experience in consulting, development, and solution architecture and delivery.
Dan Sinnreich is a Sr. Product Manager for Amazon SageMaker, focused on expanding no-code / low-code services. He is dedicated to making ML and generative AI more accessible and applying them to solve challenging problems. Outside of work, he can be found playing hockey, scuba diving, and reading science fiction.
Syed Furqhan is a Senior Software Engineer for AI and ML at AWS. He was part of many AWS service launches like Amazon Lookout for Metrics, Amazon Sagemaker and Amazon Bedrock. Currently, he is focusing on generative AI initiatives as part of Amazon Bedrock Core Systems. He is a clean code advocate and a subject-matter expert on server-less and event-driven architecture. You can follow him on linkedin, syedfurqhan
Nirmal Kumar is Sr. Product Manager for the Amazon SageMaker service. Committed to broadening access to AI/ML, he steers the development of no-code and low-code ML solutions. Outside work, he enjoys travelling and reading non-fiction.
Generative AI-powered technology operations
Technology operations (TechOps) refers to the set of processes and activities involved in managing and maintaining an organization’s IT infrastructure and services. There are several terminologies used with reference to managing information technology operations, including ITOps, SRE, AIOps, DevOps, and SysOps. For the context of this post, we refer to these terminologies as TechOps. This includes tasks such as managing servers, networks, databases, and applications to maintain reliability, performance, and security of IT systems. However, certain tasks require manual and repetitive efforts such as incident detection and response, analyzing incoming tickets from disparate service providers, finding standard operating procedures for known and unknown issues, and managing support case resolution. In recent years, TechOps has been using AI capabilities—called AIOps—for operational data collection, aggregation, and correlation to generate actionable insights, identity root causes, and more.
This post describes how AWS generative AI solutions (including Amazon Bedrock, Amazon Q Developer, and Amazon Q Business) can further enhance TechOps productivity, reduce time to resolve issues, enhance customer experience, standardize operating procedures, and augment knowledge bases. The ability of generative AI technology to interpret complex situations on a nuanced, case-by-case basis implies that generative AI can solve challenges that other approaches—including traditional artificial intelligence and machine learning (AI/ML)-based pattern matching—couldn’t handle. The following table depicts a few examples of how AWS generative AI services can help with some of the day-to-day TechOps activities.
| Amazon Bedrock | Amazon Q Developer | Amazon Q Business |
| Root cause analysis | Maintenance tasks code generation | Standard operating procedure |
| Knowledge base creation | Increase productivity and efficiency | Organization policy and procedure |
| Recurring reporting | . | Customer experience and sentiment analysis |
| Outbound support case generation | . | Shift handover chatbot |
| Inbound maintenance notifications formatting | . | . |
A typical day in the life of a TechOps team includes issue resolution, root cause analysis, maintenance activities, and updating knowledge bases to provide a positive customer experience. In the following sections, we discuss some of these areas and how generative AI can help enhance TechOps.
Event management
By monitoring systems and analyzing patterns in performance data, an AI model can predict issues before they cause outages or degraded service. When incidents do occur, generative AI models can generate preliminary documentation of the event, including details on impacted systems, potential root causes, and troubleshooting steps. This allows engineers to quickly get up to speed on new incidents and accelerate response efforts.
Generative AI can also generate summary reports of past incidents to help teams identify recurring problems and opportunities for preventative measures. Furthermore, it can help with formatting inbound maintenance notifications from various service providers into a standard format, which can speed up understanding the impact of upcoming maintenance. Similarly, generative AI can automatically generate outbound cases to service providers if it detects an anomaly.
By taking over basic documentation and prediction tasks, generative AI can help infrastructure teams spend less time on repetitive work and more time resolving issues to improve overall system reliability.
To learn more about using Amazon Bedrock for summary tasks, refer to Create summaries of recordings using generative AI with Amazon Bedrock and Amazon Transcribe. To learn how Wiz uses Amazon Bedrock to address security risks, see How Wiz is empowering organizations to remediate security risks faster with Amazon Bedrock. To learn how HappyFox uses Anthropic Claude in Amazon Bedrock, refer to HappyFox Automates Support Agent Responses with Claude in Amazon Bedrock, Increasing Ticket Resolution by 40%.
Knowledge base management
Generative AI has the potential to help engineers automatically create operational documents such as standard operating procedures (SOPs) and supplemental documents, such as server hardening, security policies for external IPs allow lists and operating system patching, and more.
Using natural language models trained on large datasets of existing SOPs and similar content, generative AI systems can understand the common structure and language used in these types of documents. Engineers can then provide the system with high-level requirements or parameters for a new procedure, and generative AI can automatically generate a draft document formatted with the appropriate sections, level of detail, and terminology. This allows engineers to spend less time on documentation and more time focused on other engineering tasks. The initial drafts from AI also provide a strong starting point that engineers can refine.
Overall, generative AI offers a more efficient way for engineers to develop standardized procedural content at scale.
To learn how to use Amazon Bedrock to generate product descriptions, see Automating product description generation with Amazon Bedrock. Additionally, refer to How Skyflow creates technical content in days using Amazon Bedrock to learn how Skyflow Inc.—a data privacy company—uses Amazon Bedrock to streamline the creation of technical content, reducing the process from weeks to days while maintaining the highest standards of data privacy and security.
Automation
Generative AI can assist engineers and automate certain tasks that would otherwise require manual work. One area this could help in is script code generation for repetitive automation processes. By training AI models on large datasets of existing code examples for common programming tasks like file operations or system configuration, generative models can learn patterns and syntax.
An Amazon Q customization is a set of elements that enables Amazon Q to provide you with suggestions based on your company’s code base. Engineers can then provide high-level descriptions or specifications of what they need automated, such as “Generate a script to back up and archive files older than 30 days in this directory.” The AI model would be able to produce working code to accomplish this automatically based on its training. This would save engineers considerable time writing and testing scripts for routine jobs, allowing them to focus on more creative and challenging aspects of their work. As generative AI techniques advance, more complex engineering automation may also be achieved.
Refer to Upgrade your Java applications with Amazon Q Code Transformation to learn about the Amazon Q Code Transformation feature. Also, refer to Using Amazon Bedrock Agents to interactively generate infrastructure as code to learn how to configure Amazon Bedrock Agents to generate infrastructure as code. Lastly, refer to TymeX Accelerates Clean Coding by 40% by Implementing Generative AI on AWS to learn how TymeX uses generative AI on AWS.
Customer experience
Generative AI can analyze large volumes of customer service data, like call logs and support tickets, and identify patterns in issues customers frequently report. This insight allows operations teams to proactively address common problems before they severely impact customers. Generative AI assistants can also automate many routine service tasks, freeing up human agents to focus on more complex inquiries that require personalization. With AI assistance, infrastructure services can be restored more quickly when outages occur. This helps make sure operations are more efficient and transparent, directly enhancing the experience for the customers that infrastructure teams aim to support.
Amazon Q Business offers a conversational experience with generative prompts and tasks that can act as a front-line support engineer, answering customer questions and resolving known issues efficiently. The feature can use data from enterprise systems to provide accurate and timely responses, reducing the burden on human engineers and improving customer satisfaction.
With Amazon Bedrock, you can perform sentiment analysis to help analyze customer emotions and provide context to human engineers, enabling them to provide better support and improve customer loyalty, retention, and growth.
Refer to Develop advanced generative AI chat-based assistants by using RAG and ReAct prompting to learn one way to develop generative AI assistants. Refer to Building a Generative AI Contact Center Solution for DoorDash Using Amazon Bedrock, Amazon Connect, and Anthropic’s Claude to learn how DoorDash built a generative AI contact center solution using AWS services. To learn how PGA TOUR built a generative AI virtual assistant, see The journey of PGA TOUR’s generative AI virtual assistant, from concept to development to prototype.
Staff productivity
An all-day infrastructure operations team faces challenges in maintaining staff productivity during off-hours and nights when the volume of support requests is lower. A generative AI assistant can help improve staff productivity in these periods and streamline the shift-handover process.
The assistant can be trained on historical support conversations to understand and resolve a large percentage of routine queries independently. It can communicate with customers on messaging platforms to provide instant assistance. Simple requests that the assistant can address free up the team to focus on complex issues requiring human expertise. The AI system can escalate any queries it can’t resolve on its own to the on-call staff. This allows the night and weekend crew to work with fewer interruptions. They can work through tasks more efficiently knowing the assistant is handling basic support needs independently. Generative AI-powered contact center solutions can improve an agent’s ability to interact with customers more precisely and speed up issue resolution, increasing overall productivity.
To learn how to automate document and data retrieval for AI assistants, see Automate chatbot for document and data retrieval using Amazon Bedrock Agents and Knowledge Bases. Refer to How LeadSquared accelerated chatbot deployments with generative AI using Amazon Bedrock and Amazon Aurora PostgreSQL to learn how LeadSquared uses Amazon Bedrock and Amazon Aurora PostgreSQL-Compatible Edition to deploy generative AI-powered assistants on their Converse platform, which personalize interactions based on customer-specific training data. This integration reduces customer onboarding costs, minimizes manual effort, and improves chatbot responses, transforming customer support and engagement by providing swift and relevant assistance.
Reporting
Generative AI has the potential to help infrastructure operations teams streamline reporting processes. By using ML algorithms trained on past report examples, a generative AI system can automatically generate draft reports based on incoming data from monitoring systems and other operational tools. This can save teams significant time spent compiling information into standardized report formats. The AI-generated reports could include summary data visualizations, descriptive analyses, and recommendations tailored to each recipient.
Teams would still need to review the drafts for accuracy before finalizing and distributing them. However, having an initial version generated automatically could cut down on routine reporting tasks so engineers have more time for higher-value problem-solving and strategic planning work. The use of AI could help infrastructure teams meet their reporting obligations more efficiently.
Amazon Q in QuickSight is your generative AI assistant that makes it straightforward to build and consume insights. For more information, see Amazon Q is now generally available in Amazon QuickSight, bringing Generative BI capabilities to the entire organization. Also, refer to Anthology uses embedded analytics offered by Amazon QuickSight to democratize decision making for higher education to learn how Anthology is using Amazon Q in QuickSight to offer institutions self-serve options for analytics needs that aren’t directly addressed by the central dashboards.
You can explore more customer stories and case studies at Generative AI Customer Stories to learn how customers are using AWS generative AI services. Refer to Derive meaningful and actionable operational insights from AWS Using Amazon Q Business to learn how to use AWS generative AI services, like Amazon Q Business, with AWS Support cases, AWS Trusted Advisor, and AWS Health data to derive actionable insights based on common patterns, issues, and resolutions while using the AWS recommendations and best practices enabled by support data.
Conclusion
Integrating generative AI into TechOps represents a transformative leap in the management and optimization of IT infrastructure and services. By using AWS generative AI solutions such as Amazon Bedrock, Amazon Q Developer, and Amazon Q Business, organizations can significantly enhance productivity, reduce the time required to resolve issues, and improve overall customer experience. Generative AI’s sophisticated capabilities in predicting and preventing outages, automating documentation, and generating actionable insights from operational data position it as a critical tool for modern TechOps teams.
You can unlock unimagined possibilities with generative AI by using the AWS Generative AI Innovation Center program, which pairs you with AWS science and strategy experts with deep experience in AI/ML and generative AI techniques. To get started, contact your AWS Account Manager. If you don’t have an AWS Account Manager, contact AWS Sales.
About the Authors
Raman Pujani is a Solutions Architect at Amazon Web Services, where he helps customers to accelerate their business transformation journey with AWS. He builds simplified and sustainable solutions for complex business problems with innovative technology. Raman has 25+ years of experience in IT Transformation. Besides work, he enjoys spending time with family, vacationing in the mountains, and music.
Rachanee Singprasong is a Principal Customer Solutions Manager in Strategic Accounts at Amazon Web Services. Her role is focused on enabling customer in their cloud adoption and digital transformation journey. She has a Ph.D. in Operations Research and her passion is to solve complex customer challenges using creative solutions.
Vijay Sivaji is a Senior Technical Account Manager in Strategic Accounts at Amazon Web Services. He helps customers in solving architectural, operational and cost optimization challenges. In his spare time he enjoys playing tennis.
Optimizing MLOps for Sustainability
Machine learning operations (MLOps) are a set of practices that automate and simplify machine learning (ML) workflows and deployments. What is MLOps provides a detailed description of this concept. As ML workloads become increasingly complex and consume more energy and resources, a growing number of companies are looking for ways to manage both the costs and the carbon footprint associated with these workloads. AWS published Guidance for Optimizing MLOps for Sustainability on AWS to help customers maximize utilization and minimize waste in their ML workloads.
In this blog post, you will learn how to optimize MLOps for sustainability.
There are three main workflows in the overall process for building, deploying and using ML models, as shown in the following figure. The process begins with data preparation, followed by model training and tuning, and then model deployment and management.
Data preparation
The workflow starts with data preparation, which includes four components: your data stream, Amazon SageMaker Processing job, Amazon SageMaker Feature Store and an Amazon Simple Storage Service (Amazon S3) bucket for raw data, as shown in the following figure.
Data preparation is essential for model training and is also the first phase in the MLOps lifecycle. Optimizing the artificial intelligence and machine learning (AI/ML) data preparation workload on AWS with sustainability best practices helps reduce the carbon footprint and the cost.
The data preparation process can be complex and energy-intensive because of the vast amount of data processing and computations involved. This leads to substantial resource consumption. There are a few things to consider that can help reduce energy consumption.
Start with the AWS Region you choose for your workload. If possible, choose a Region that has low carbon intensity or where the electricity is attributed to 100% renewable energy sources. In addition, consider storing data and training models in the same Region if possible. This reduces the data movement and latency across the network, optimizing the networking resources required.
Using a serverless architecture can help further reduce resource consumption and remove maintenance overhead by provisioning resources only when required. It’s also important to avoid duplication and re-run of code across teams. Look for services such as Amazon SageMaker Feature Store which helps achieve this goal. Finally, choosing the right storage type for the data used for model training can limit the carbon impact of your workload.
For example, by using S3 One Zone-Infrequent Access to store data that isn’t frequently accessed, such as test data and training data, you can optimize the carbon impact of the data stored. Also, using S3 Intelligent-Tiering can help move the data to more energy-efficient tiers based on access patterns.
Model training and tuning
The second area for you to consider is model training and tuning, shown in the following figure.
While data preparation isn’t unique to AI/ML workloads, the model training and tuning workflow is specific to AI/ML. It’s an important step in making the models functionally useful while also reducing the resources required to run them at scale. There are costs in terms of both operations and sustainability. The good news is that optimization for sustainability also helps to optimizing operations.
For example, SageMaker provides the model parallel library to help efficiently distribute and train models on multiple compute nodes. The library has multiple features that can be combined to more efficiently train models from relatively small parameter sets up to sets with hundreds of billions of parameters. The library can also help use the features of Elastic Fabric Adapter (EFA) supported devices to maximize throughput and minimize latency across nodes. Further optimization is possible using SageMaker Training Compiler to compile deep learning models for training on supported GPU instances. SageMaker Training Compiler converts deep learning models from high-level language representation to hardware-optimized instructions. Hardware-optimized instructions can speed up model training by up to 50% by more efficiently using the GPU memory and using a larger batch size per iteration, all without altering the final trained model.
To reduce the time and energy required to tune a model, SageMaker automatic model tuning (AMT) runs multiple training jobs on a given dataset; it then uses the results to converge on a set of hyperparameter values to create the best performing model for a given metric. There are multiple approaches to the process of searching for the right hyperparameter ranges. For example, Bayesian optimization typically requires 10 times fewer jobs to find the best set of values compared to other methods, reducing the resource usage and carbon footprint of the process.
Right-sizing is another method for managing resource usage and minimizing the environmental impact of your workloads. SageMaker debugger helps to optimize resource consumption by detecting under-utilization of system resources, identifying training problems, and using built-in rules to monitor and stop training jobs as soon as bugs are detected.
Data pre- and post-processing and model evaluation tasks can be run as Amazon SageMaker Processing jobs. In addition to evaluating the accuracy of your models, processing jobs help you to make informed decisions about the tradeoffs between a model’s accuracy and its carbon footprint. Thus, you can establish performance criteria that support your sustainability goals while meeting your business requirements. SageMaker Processing also provides Amazon CloudWatch logs and metrics that can be used for monitoring and right-sizing jobs based on CPU, memory, GPU, GPU memory, and disk metrics.
Dedicated Amazon Elastic Compute Cloud (Amazon EC2) Trn1 instances provide both efficiency and environmental benefits for running your training jobs. These instances use Trainium processors: purpose-built chips designed specifically for deep learning training of models that can exceed 100 billion parameters. Each Trn1 instance provides up to 16 Trainium accelerators, ensuring that jobs will be both efficient and cost optimized. EC2 Trn1 instances offer up to 52% cost-to-train savings compared to comparable EC2 instance types.
Next, you can use governance to share information about the environmental impact of your model. Amazon SageMaker Model Cards provide versioned records documenting various aspects and attributes of your model. This allows you to share the intended uses and assessed carbon impact of a model so that data scientists, ML engineers, and other teams can make informed decisions when choosing and running models.
Model deployment and management
The last area of MLOps is deployment and management, shown in the following figure.
Automating the deployment of ML models provides several sustainability benefits. The deployed model can use a lot of resources when data or code is updated and retrained. You want to ensure that the deployed model is as efficient as possible to reduce the carbon footprint of the workload.
One approach is to use Amazon SageMaker Model Registry. This feature helps improve sustainability and resource optimization by providing a centralized repository for cataloging ML models and reducing redundancy. This approach improves model reusability by allowing existing models to be fine-tuned, rather than training new models from scratch. Consider running your deployment code using AWS CodePipeline to ensure repeatability and version control and optimize resource utilization by running only the necessary stages in the pipeline. This helps your workloads remove the waste associated with manual processes and supports incremental improvements over time.
If your workloads can tolerate latency, consider deploying your model on Amazon SageMaker Asynchronous Inference with auto-scaling groups. This can help minimize idle resources and reduce the impact of load spikes. This also means you pay for compute only when the endpoint is actively handling inference requests. Alternatively, if you don’t need real-time inference, use batch transform. Unlike persistent endpoints, clusters are decommissioned when a batch transform job is complete. Batch transform automatically partitions large datasets and distributes workloads across compute to ensure efficient resource utilization.
To simplify deployment and management and increase resource utilization, use multi-model endpoints instead of separate endpoints for each model. One example for this approach is models with different data formats, such as recommendation systems that process text and images using separate endpoints. Or deploying a variety of models that include PyTorch, Scikit-learn, and TensorFlow models. Automatic scaling can amplify resource optimization for your hosted models. Auto scaling dynamically adjusts the number of instances provisioned for a model in response to changes in your workload. This helps you avoid cost and consumes less energy and resources. If your workload has intermittent or unpredictable traffic with idle periods between traffic peaks and can tolerate cold starts, use Amazon SageMaker Serverless Inference endpoints, which automatically launch compute resources and scale depending on traffic. Optionally, you can use Provisioned Concurrency with Serverless Inference when you have predictable bursts in your traffic.
AWS offers a few different options to better utilize your resources and lower emissions when working with inference workloads. AWS Inferentia is designed to deliver high performance at the lowest cost in EC2 instances for your deep learning and generative AI inference applications. AWS Inferentia is built for sustainability and provides up to 50% better performance per watt over comparable EC2 instances. You can further optimize resource utilization by combining AWS Inferentia and Amazon Elastic Inference to attach the right amount of GPU-powered inference acceleration to any EC2 or SageMaker instance type.
After training a model for high accuracy, developers often turn to more expensive large instances with lots of memory and processing power to achieve better throughput. You can reduce resource usage and avoid the need for more powerful instances by using pre-trained models and compiling them into optimized executables that can be hosted in SageMaker or edge devices for inference with Amazon SageMaker Neo.
Monitoring CPU, memory, and GPU resource utilization is critical to optimize model performance and avoid wasted resources. AWS offers a variety of tools that you can use to optimize MLOps for sustainability, such as CloudWatch, SageMaker Inference recommender, and SageMaker Model Monitor. Inference Recommender helps you choose the optimal instance type and configuration for ML models and workloads. You can use SageMaker Model Monitor to automate drift detection of your ML model in production, and only retrain it when prediction performance drops below predetermined key performance indicators (KPIs). This approach improves operational efficiency and retrains the model based on your business metrics.
Conclusion
Sustainability and ML are redefining how many companies deliver value for their customers. Incorporating sustainability into the design, development and deployment of ML models is a crucial long-term consideration. AWS is investing in the sustainability of the cloud and providing resources to assist customers in transforming their workloads to be more energy efficient. In this post, we have reviewed the Guidance for Optimizing MLOps for Sustainability on AWS, providing service-specific practices to understand and reduce the environmental impact of these workloads. MLOps consists of several distinct phases that can be independently optimized for sustainability. Regular reviews using tools such as AWS Well-Architected Machine Learning Lens help you identify optimization opportunities and provide a mechanism for you to meet your sustainability goals.
About the Authors
Archana Srinivasan is a Senior Technical Account Manager within Enterprise Support at Amazon Web Services (AWS). Archana provides strategic technical guidance for independent software vendors (ISVs) to innovate and operate their workloads efficiently on AWS.
Chris Procunier is a Senior Technical Account Manager at AWS, based out of Washington DC. He has been managing systems and infrastructure for 25 years as an entrepreneur, IT Director and architect. Outside of work Chris is passionate about family, friends, music, cooking and cycling.
Meghana Reddy is a Technical Account Manager at AWS, where she offers strategic technical guidance to Independent Software Vendors (ISVs) for optimizing their workloads on AWS. She is passionate about environmental sustainability and actively promotes sustainable practices within the cloud.
Steven David is a Principal Solutions Architect at Amazon Web Services (AWS). He has over 20 years of experience designing solutions for large enterprises. Through these engagements he has developed deep expertise in application development technologies and methodologies.
NotebookLM now lets you listen to a conversation about your sources
NotebookLM is releasing Audio Overviews, which turns your sources into an engaging discussion.Read More
Enabling complex generative AI applications with Amazon Bedrock Agents
In June, I started a series of posts that highlight the key factors that are driving customers to choose Amazon Bedrock. The first covered building generative AI apps securely with Amazon Bedrock, while the second explored building custom generative AI applications with Amazon Bedrock. Now I’d like to take a closer look at Amazon Bedrock Agents, which empowers our customers to build intelligent and context-aware generative AI applications, streamlining complex workflows and delivering natural, conversational user experiences. The advent of large language models (LLMs) has enabled humans to interact with computers using natural language. However, many real-world scenarios demand more than just language comprehension. They involve executing complex multi-step workflows, integrating external data sources, or seamlessly orchestrating diverse AI capabilities and data workflows. In these real-world scenarios, agents can be a game changer, delivering more customized generative AI applications—and transforming the way we interact with and use LLMs.
Answering more complex queries
Amazon Bedrock Agents enables a developer to take a holistic approach in improving scalability, latency, and performance when building generative AI applications. Generative AI solutions that use Amazon Bedrock Agents can handle complex tasks by combining an LLM with other tools. For example, imagine that you are trying to create a generative AI-enabled assistant that helps people plan their vacations. You want it to be able to handle simple questions like “What’s the weather like in Paris next week?” or “How much does it cost to fly to Tokyo in July?” A basic virtual assistant might be able to answer those questions drawing from preprogrammed responses or by searching the Internet. But what if someone asks a more complicated question, like “I want to plan a trip to three countries next summer. Can you suggest a travel itinerary that includes visiting historic landmarks, trying local cuisine, and staying within a budget of $3,000?” That is a harder question because it involves planning, budgeting, and finding information about different destinations.
Using Amazon Bedrock Agents, a developer can quickly build a generative assistant to help answer this more complicated question by combining the LLM’s reasoning with additional tools and resources, such as natively integrated knowledge bases to propose personalized itineraries. It could search for flights, hotels, and tourist attractions by querying travel APIs, and use private data, public information for destinations, and weather—while keeping track of the budget and the traveler’s preferences. To build this agent, you would need an LLM to understand and respond to questions. But you would also need other modules for planning, budgeting, and accessing travel information.
Agents in action
Our customers are using Amazon Bedrock Agents to build agents—and agent-driven applications—quickly and effectively. Consider Rocket, the fintech company that helps people achieve home ownership and financial freedom:
“Rocket is poised to revolutionize the homeownership journey with AI technology, and agentic AI frameworks are key to our mission. By collaborating with AWS and leveraging Amazon Bedrock Agents, we are enhancing the speed, accuracy, and personalization of our technology-driven communication with clients. This integration, powered by Rocket’s 10 petabytes of data and industry expertise, ensures our clients can navigate complex financial moments with confidence.”
– Shawn Malhotra, CTO of Rocket Companies.
A closer look at how agents work
Unlike LLMs that provide simple lookup or content-generation capabilities, agents integrate various components with an LLM to create an intelligent orchestrator capable of handling sophisticated use cases with nuanced context and specific domain expertise. The following figure outlines the key components of Amazon Bedrock Agents:
The process starts with two parts—the LLM and the orchestration prompt. The LLM—often implemented using models like those in the Anthropic Claude family or Meta Llama models—provides the basic reasoning capabilities. The orchestration prompt is a set of prompts or instructions that guide the LLM when driving the decision-making process.
In the following sections, we discuss the key components of Amazon Bedrock Agents in depth:
Planning: A path from task to goals
The planning component for LLMs entails comprehending tasks and devising multi-step strategies to address a problem and fulfill the user’s need. In Amazon Bedrock Agents, we use chain-of-thought prompting in combination with ReAct in the orchestration prompt to improve an agent’s ability to solve multi-step tasks. In task decomposition, the agent must understand the intricacies of an abstract request. Continuing to explore our travel scenario, if a user wants to book a trip, the agent must recognize that it encompasses transportation, accommodation, reservations for sightseeing attractions, and restaurants. This ability to split up an abstract request, such as planning a trip, into detailed, executable actions, is the essence of planning. However, planning extends beyond the initial formulation of a plan, because during execution, the plan may get dynamically updated. For example, when the agent has completed arranging transportation and progresses to booking accommodation, it may encounter circumstances where no suitable lodging options align with the original arrival date. In such scenarios, the agent must determine whether to broaden the hotel search or revisit alternative booking dates, adapting the plan as it evolves.
Memory: Home for critical information
Agents have both long-term and short-term memory. Short-term memory is detailed and exact. It is relevant to the current conversation and resets when the conversation is over. Long-term memory is episodic and remembers important facts and details in the form of saved summaries. These summaries serve as the memory synopses of previous dialogues. The agent uses this information from the memory store to better solve the current task. The memory store is separate from the LLM, with a dedicated storage and a retrieval component. Developers have the option to customize and control which information is stored (or excluded) in memory. An identity management feature, which associates memory with specific end-users, gives developers the freedom to identify and manage end-users—and enable further development on top of Amazon Bedrock agents’ memory capabilities. The industry-leading memory retention functionality of Amazon Bedrock—launched at the recent AWS New York Summit—allows agents to learn and adapt to each user’s preferences over time, enabling more personalized and efficient experiences across multiple sessions for the same user. It is straightforward to use, allowing users to get started in a single click.
Communication: Using multiple agents for greater efficiency and effectiveness
Drawing from the powerful combination of the capabilities we’ve described, Amazon Bedrock Agents makes it effortless to build agents that transform one-shot query responders into sophisticated orchestrators capable of tackling complex, multifaceted use cases with remarkable efficiency and adaptability. But what about using multiple agents? LLM-based AI agents can collaborate with one another to improve efficiency in solving complex questions. Today, Amazon Bedrock makes it straightforward for developers to connect them through LangGraph, part of LangChain, the popular open source tool set. The integration of LangGraph into Amazon Bedrock empowers users to take advantage of the strengths of multiple agents seamlessly, fostering a collaborative environment that enhances the overall efficiency and effectiveness of LLM-based systems.
Tool Integration: New tools mean new capabilities
New generations of models such as Anthropic Claude Sonnet 3.5, Meta Llama 3.1, or Amazon Titan Text Premier are better equipped to use reources. Using these resources requires that developers keep up with ongoing updates and changes, requiring new prompts every time. To reduce this burden, Amazon Bedrock simplifies interfacing with different models, making it effortless to take advantage of all the features a model has to offer. For example, the new code interpretation capability announced at the recent AWS New York Summit allows Amazon Bedrock agents to dynamically generate and run code snippets within a secure, sandboxed environment to address complex tasks like data analysis, visualization, text processing, and equation solving. It also enables agents to process input files in various formats—including CSV, Excel, JSON—and generate charts from data.
Guardrails: Building securely
Accuracy is critical when dealing with complex queries. Developers can enable Amazon Bedrock Guardrails to help reduce inaccuracies. Guardrails improve the behavior of the applications you’re building, increase accuracy, and help you build responsibly. They can prevent both malicious intent from users and potentially toxic content generated by AI, providing a higher level of safety and privacy protection.
Amplifying and extending the capabilities of generative AI with Amazon Bedrock Agents
Enterprises, startups, ISVs, and systems integrators can take advantage of Amazon Bedrock Agents today because it provides development teams with a comprehensive solution for building and deploying AI applications that can handle complex queries, use private data sources, and adhere to responsible AI practices. Developers can start with tested examples—so-called golden utterances (input prompts) and golden responses (expected outputs). You can continuously evolve agents to fit your top use cases and kickstart your generative AI application development. Agents unlock significant new opportunities to build generative AI applications to truly transform your business. It will be fascinating to see the solutions—and results—that Amazon Bedrock Agents inspires.
Resources
For more information on customization with Amazon Bedrock, see the following resources:
- Learn about Amazon Bedrock
- Learn more about Amazon Bedrock Agents
- Learn how to associate a guardrail with your agent
- Read the announcement post on code interpretation and memory retention in Amazon Bedrock Agents
- Get hands-on insights on using Amazon Bedrock Agents
About the author
Vasi Philomin is VP of Generative AI at AWS. He leads generative AI efforts, including Amazon Bedrock and Amazon Titan.
Ready to Roll: Nuro to License Its Autonomous Driving System
To accelerate autonomous vehicle development and deployment timelines, Nuro announced today it will license its Nuro Driver autonomous driving system directly to automakers and mobility providers.
The Nuro Driver is built on NVIDIA’s end-to-end safety architecture, which includes NVIDIA GPUs for AI training in the cloud and an automotive-grade NVIDIA DRIVE Thor computer running the NVIDIA DriveOS operating system inside the vehicle.
The Nuro Driver has demonstrated its reliability and safety in real-world conditions with more than 1 million autonomous miles completed across its fleet of R&D vehicles and zero at-fault incidents.
“It’s not a question of if, but when L4 autonomy will become widespread,” said Jiajun Zhu, cofounder and CEO at Nuro. “We believe Nuro is positioned to be a major contributor to this autonomous future where people and goods mobility are free-flowing, representing a significant increase in the quality of life for everyone.”
The licensing of the Nuro Driver marks a significant step forward in bringing level 4 vehicles to market, accelerating the adoption of autonomous technology across the transportation industry.
An End-to-End Approach With NVIDIA DRIVE
Nuro announced at GTC in March that the Nuro Driver, which enables level 4 autonomous driving for multiple vehicle types, is being built on NVIDIA DRIVE Thor, running on the NVIDIA DriveOS operating system for safe, AI-defined autonomous vehicles.
DRIVE Thor integrates the NVIDIA Blackwell architecture, designed for transformer, large language models and generative AI workloads. Nuro also uses NVIDIA GPUs for AI training.
“Built with NVIDIA’s end-to-end safety AV architecture, the Nuro Driver can integrate sensor processing and other safety-critical capabilities, along with AI-driven autonomy, into a single, centralized computing system,” said Rishi Dhall, vice president of automotive at NVIDIA. “This enables the reliability and performance needed for safe deployment of autonomous vehicles at scale.”
The next-gen Nuro Driver will include safety features such as microphones for siren detection and systems for removing dirt from sensors as well as redundancy in safety-critical systems.
Advantages of Licensing
Nuro’s licensing model will offer automotive manufacturers and mobility companies access to a commercially independent, road-proven platform that can accelerate their autonomous vehicle development and deployment timelines.
With a focus on advancing autonomy, Nuro is poised to help shape the future of transportation by driving industry-wide adoption and commercialization of autonomous technology across a broad range of vehicles and mobility applications.
Test Area Expansion
Nuro this summer received approval from the California Department of Motor Vehicles to test its driverless vehicles based on the Nuro Driver in four San Francisco Bay Area cities: Los Altos, Menlo Park, Mountain View and Palo Alto.
The DMV permit allows Nuro vehicles to travel at any time of the day, as well as in light rain and light to moderate fog conditions.
Nuro is also conducting commercial testing and delivery services in Houston.
NVIDIA and Oracle to Accelerate AI and Data Processing for Enterprises
Enterprises are looking for increasingly powerful compute to support their AI workloads and accelerate data processing. The efficiency gained can translate to better returns for their investments in AI training and fine-tuning, and improved user experiences for AI inference.
At the Oracle CloudWorld conference today, Oracle Cloud Infrastructure (OCI) announced the first zettascale OCI Supercluster, accelerated by the NVIDIA Blackwell platform, to help enterprises train and deploy next-generation AI models using more than 100,000 of NVIDIA’s latest-generation GPUs.
OCI Superclusters allow customers to choose from a wide range of NVIDIA GPUs and deploy them anywhere: on premises, public cloud and sovereign cloud. Set for availability in the first half of next year, the Blackwell-based systems can scale up to 131,072 Blackwell GPUs with NVIDIA ConnectX-7 NICs for RoCEv2 or NVIDIA Quantum-2 InfiniBand networking to deliver an astounding 2.4 zettaflops of peak AI compute to the cloud. (Read the press release to learn more about OCI Superclusters.)
At the show, Oracle also previewed NVIDIA GB200 NVL72 liquid-cooled bare-metal instances to help power generative AI applications. The instances are capable of large-scale training with Quantum-2 InfiniBand and real-time inference of trillion-parameter models within the expanded 72-GPU NVIDIA NVLink domain, which can act as a single, massive GPU.
This year, OCI will offer NVIDIA HGX H200 — connecting eight NVIDIA H200 Tensor Core GPUs in a single bare-metal instance via NVLink and NVLink Switch, and scaling to 65,536 H200 GPUs with NVIDIA ConnectX-7 NICs over RoCEv2 cluster networking. The instance is available to order for customers looking to deliver real-time inference at scale and accelerate their training workloads. (Read a blog on OCI Superclusters with NVIDIA B200, GB200 and H200 GPUs.)
OCI also announced general availability of NVIDIA L40S GPU-accelerated instances for midrange AI workloads, NVIDIA Omniverse and visualization. (Read a blog on OCI Superclusters with NVIDIA L40S GPUs.)
For single-node to multi-rack solutions, Oracle’s edge offerings provide scalable AI at the edge accelerated by NVIDIA GPUs, even in disconnected and remote locations. For example, smaller-scale deployments with Oracle’s Roving Edge Device v2 will now support up to three NVIDIA L4 Tensor Core GPUs.
Companies are using NVIDIA-powered OCI Superclusters to drive AI innovation. Foundation model startup Reka, for example, is using the clusters to develop advanced multimodal AI models to develop enterprise agents.
“Reka’s multimodal AI models, built with OCI and NVIDIA technology, empower next-generation enterprise agents that can read, see, hear and speak to make sense of our complex world,” said Dani Yogatama, cofounder and CEO of Reka. “With NVIDIA GPU-accelerated infrastructure, we can handle very large models and extensive contexts with ease, all while enabling dense and sparse training to scale efficiently at cluster levels.”
Accelerating Generative AI Oracle Database Workloads
Oracle Autonomous Database is gaining NVIDIA GPU support for Oracle Machine Learning notebooks to allow customers to accelerate their data processing workloads on Oracle Autonomous Database.
At Oracle CloudWorld, NVIDIA and Oracle are partnering to demonstrate three capabilities that show how the NVIDIA accelerated computing platform could be used today or in the future to accelerate key components of generative AI retrieval-augmented generation pipelines.
The first will showcase how NVIDIA GPUs can be used to accelerate bulk vector embeddings directly from within Oracle Autonomous Database Serverless to efficiently bring enterprise data closer to AI. These vectors can be searched using Oracle Database 23ai’s AI Vector Search.
The second demonstration will showcase a proof-of-concept prototype that uses NVIDIA GPUs, NVIDIA RAPIDS cuVS and an Oracle-developed offload framework to accelerate vector graph index generation, which significantly reduces the time needed to build indexes for efficient vector searches.
The third demonstration illustrates how NVIDIA NIM, a set of easy-to-use inference microservices, can boost generative AI performance for text generation and translation use cases across a range of model sizes and concurrency levels.
Together, these new Oracle Database capabilities and demonstrations highlight how NVIDIA GPUs can be used to help enterprises bring generative AI to their structured and unstructured data housed in or managed by an Oracle Database.
Sovereign AI Worldwide
NVIDIA and Oracle are collaborating to deliver sovereign AI infrastructure worldwide, helping address the data residency needs of governments and enterprises.
Brazil-based startup Wide Labs trained and deployed Amazonia IA, one of the first large language models for Brazilian Portuguese, using NVIDIA H100 Tensor Core GPUs and the NVIDIA NeMo framework in OCI’s Brazilian data centers to help ensure data sovereignty.
“Developing a sovereign LLM allows us to offer clients a service that processes their data within Brazilian borders, giving Amazônia a unique market position,” said Nelson Leoni, CEO of Wide Labs. “Using the NVIDIA NeMo framework, we successfully trained Amazônia IA.”
In Japan, Nomura Research Institute, a leading global provider of consulting services and system solutions, is using OCI’s Alloy infrastructure with NVIDIA GPUs to enhance its financial AI platform with LLMs operating in accordance with financial regulations and data sovereignty requirements.
Communication and collaboration company Zoom will be using NVIDIA GPUs in OCI’s Saudi Arabian data centers to help support compliance with local data requirements.
And geospatial modeling company RSS-Hydro is demonstrating how its flood mapping platform — built on the NVIDIA Omniverse platform and powered by L40S GPUs on OCI — can use digital twins to simulate flood impacts in Japan’s Kumamoto region, helping mitigate the impact of climate change.
These customers are among numerous nations and organizations building and deploying domestic AI applications powered by NVIDIA and OCI, driving economic resilience through sovereign AI infrastructure.
Enterprise-Ready AI With NVIDIA and Oracle
Enterprises can accelerate task automation on OCI by deploying NVIDIA software such as NIM microservices and NVIDIA cuOpt with OCI’s scalable cloud solutions. These solutions enable enterprises to quickly adopt generative AI and build agentic workflows for complex tasks like code generation and route optimization.
NVIDIA cuOpt, NIM, RAPIDS and more are included in the NVIDIA AI Enterprise software platform, available on the Oracle Cloud Marketplace.
Learn More at Oracle CloudWorld
Join NVIDIA at Oracle CloudWorld 2024 to learn how the companies’ collaboration is bringing AI and accelerated data processing to the world’s organizations.
Register to the event to watch sessions, see demos and join Oracle and NVIDIA for the solution keynote, “Unlock AI Performance with NVIDIA’s Accelerated Computing Platform” (SOL3866), on Wednesday, Sept. 11, in Las Vegas.