Novel view synthesis from a single image requires inferring occluded regions of objects and scenes while simultaneously maintaining semantic and physical consistency with the input. Existing approaches condition neural radiance fields (NeRF) on local image features, projecting points to the input image plane, and aggregating 2D features to perform volume rendering. However, under severe occlusion, this projection fails to resolve uncertainty, resulting in blurry renderings that lack details. In this work, we propose NerfDiff, which addresses this issue by distilling the knowledge of a 3D-aware…Apple Machine Learning Research
Unlock Insights from your Amazon S3 data with intelligent search
Amazon Kendra is an intelligent search service powered by machine learning (ML). Amazon Kendra reimagines enterprise search for your websites and applications so your employees and customers can easily find the content they’re looking for, even when it’s scattered across multiple locations and content repositories within your organization. Keywords or natural language questions can be used to search most relevant documents powered by ML to deliver answers and rank documents. Amazon Kendra can index data from Amazon Simple Storage Service (Amazon S3) or from a third-party document repository. Amazon S3 is an object storage service that offers scalability and availability where you can store large amounts of data, including product manuals, project and research documents, and more.
In this post, you can learn how to deploy a provided AWS CloudFormation template to index your documents in an Amazon S3 bucket. The template creates an Amazon Kendra data source for an index and synchronizes your data source according to your needs: on-demand, hourly, daily, weekly or monthly. AWS CloudFormation allows us to provision infrastructure as code (IaC) so you can spend less time managing resources, replicate your infrastructure quickly, and control and track changes in the infrastructure.
Overview of the solution
The CloudFormation template sets up an Amazon Kendra data source with a connection to Amazon S3. The template also creates one role for the Amazon Kendra data source service. You can specify an S3 bucket, synchronization schedule, and inclusion/exclusion patterns. When the synchronization job has finished, you can search the indexed content through the Search console. The following diagram illustrates this workflow.
This post guides you to the following steps:
- Deploy the provided template.
- Upload the documents to the S3 bucket that you create. If you provide a bucket with documents, you can omit this step.
- Wait until the index finishes crawling the data source.
Prerequisites
For this walkthrough, you should have the following prerequisites:
- An AWS account where the proposed solution can be deployed.
- An Amazon Kendra index for attaching a data source to the stack.
- The set of documents that are used to create the Amazon Kendra index. In this solution, you are using a compressed file of AWS whitepapers.
Deploy the solution with AWS CloudFormation
To deploy the CloudFormation template, complete the following steps:
- Choose
You’re redirected to the AWS CloudFormation console.
- You can modify the parameters or use the default values:
- The Amazon Kendra data source name is automatically set using the stack name and associated bucket name.
- For KendraIndexId, enter the Amazon Kendra index ID where you will attach the data source.
- You can also choose when you want to run the data source synchronization using KendraSyncSchedule. By default, it’s set to OnDemand.
- For S3BucketName, you can either enter a bucket you have already created or leave it empty. If you leave it empty, a bucket will be created for you. Either way, the bucket is used as the Amazon Kendra data source. For this post, we leave it empty.
It takes around 5 minutes for the stack to deploy the Amazon Kendra data source attached to the Amazon Kendra index.
- On the Outputs tab of the CloudFormation stack, copy the name of the created bucket, data source name, and ID.
The created stack deploys one role: <stack-name>-KendraDataSourceRole. It’s a best practice to deploy a role for each data source you create. This role gives Amazon Kendra data source to add or remove files from Amazon Kendra index, to get objects from Amazon S3 bucket.
Upload files to the S3 bucket
Amazon Kendra can handle multiple document types, such as .html, .pdf, .csv, .json, .docx, and .ppt. You can also have a combination of documents on a single index. The text contained in those documents is indexed to the provided Amazon Kendra index. You can search for keywords on AWS topics on best practices, databases, machine learning, security, and more using over 60 pdf files that you can download. For example, if you want to know where you can find more information about caching in the AWS whitepapers, Amazon Kendra can help you find documents related to databases and best practices.
When you download the AWS Whitepapers.zip file and uncompress the file, you see these six folders: Best_Practices, Databases, General, Machine_Learning, Security, Well_Architected. Upload these folders to your S3 bucket.
Synchronize the Amazon Kendra data source
Amazon Kendra data source data can synchronize your data based on preconfigured schedule or can be be manually triggered on-demand. By default, CloudFormation template configures the data source to on-demand synchronization schedule to be triggered manually as required.
To manually trigger the synchronization job from the AWS Amazon Kendra console, navigate to the Amazon Kendra index used as part of CloudFormation stack deployment, under Data Management in the navigation pane, choose Data Sources and then choose Sync now. This makes the S3 bucket synchronize with the data source.
When the Amazon Kendra data source starts syncing, you should see the Current sync state as Syncing.
When the data source has finished, the Last sync status appears as Succeeded and Current sync state as Idle. You can now search the indexed content.
Configure synchronization schedule
The template allows you to run the schedule every hour at minute 0, for example, 13:00, 14:00, or 15:00. You also have the option to run it daily at 00:00 UTC. The Weekly setting runs Mondays at 00:00 UTC, and the Monthly setting runs every first day of the month at 00:00 UTC.
To change the schedule after the Amazon Kendra data source has been created, on the Actions menu, choose Edit. Under Configure sync settings, you find the Sync rule schedule section.
Under Frequency, you can select hourly, daily, weekly, monthly, or custom, all of which allow you to schedule your sync down to the minute.
Add exclusion patterns
The provided CloudFormation template allows you to add exclusion patterns. By default, .png and .jpg files will be added to the ExclusionPatterns parameter. Additional file formats can be added as a comma separated list to the exclusion pattern. Similarly, InclusionPatterns parameter may be used add comma list file formats to set up an inclusion pattern. If you don’t provide an inclusion pattern, all files are indexed except for the ones included in the exclusion parameter.
Clean up
To avoid costs, you can delete the stack from the AWS CloudFormation console. On the Stacks page, select the stack you created, choose Delete, and confirm the deletion of the stack.
If you haven’t provided a S3 bucket, the stack creates a bucket. If the bucket is empty, it’s automatically deleted. Otherwise, you need to empty the folder and manually delete it. If you provided a bucket, even if it’s empty, it won’t be deleted. Amazon Kendra index won’t be deleted. Only the Amazon Kendra data source created by the stack will be deleted.
Conclusion
In this post, we provided an CloudFormation template to easily synchronize your text documents on an S3 bucket to your Amazon Kendra index. This solution is helpful if you have multiple S3 buckets you want to index because you can create all the necessary components to query the documents with a few clicks in a consistent and repeatable manner. You can also see how image-based text documents can be handled in Amazon Kendra. To learn more about specific schedule patterns, refer to Schedule Expressions for Rules.
Leave a comment and learn more about Amazon Kendra index creation in the following Amazon Kendra Essentials+ workshop.
Special thanks to Jose Mauricio Mani Yanez for his help creating the example code and compiling the content for this post.
About the author
Rajesh Kumar Ravi is an AI/ML Specialist Solutions Architect at Amazon Web Services specializing in intelligent document search with Amazon Kendra and generative AI. He is a builder and problem solver, and contributes to development of new ideas. He enjoys walking and loves to go on short hiking trips outside of work.
Language Identification: Building an End-to-End AI Solution using PyTorch
Language Identification is the process of identifying the primary language from multiple audio input samples. In natural language processing (NLP), language identification is an important problem and a challenging issue. There are many language-related tasks such as entering text on your phone, finding news articles you enjoy, or discovering answers to questions that you may have. All these tasks are powered by NLP models. To decide which model to invoke at a particular point in time, we must perform language identification.
This article presents an in-depth solution and code sample for language identification using Intel® Extension for PyTorch, which is a version of the popular PyTorch AI framework optimized for use on Intel® processors, and Intel® Neural Compressor, which is a tool to accelerate AI inference without sacrificing accuracy.
The code sample demonstrates how to train a model to perform language identification using the Hugging Face SpeechBrain* toolkit and optimize it using the Intel® AI Analytics Toolkit (AI Kit). The user can modify the code sample and identify up to 93 languages using the Common Voice dataset.
Proposed Methodology for Language Identification
In the proposed solution, the user will use an Intel AI Analytics Toolkit container environment to train a model and perform inference leveraging Intel-optimized libraries for PyTorch. There is also an option to quantize the trained model with Intel Neural Compressor to speed up inference.
Dataset
The Common Voice dataset is used and for this code sample, specifically, Common Voice Corpus 11.0 for Japanese and Swedish. This dataset is used to train an Emphasized Channel Attention, Propagation and Aggregation Time Delay Neural Network (ECAPA-TDNN), which is implemented using the Hugging Face SpeechBrain library. Time Delay Neural Networks (TDNNs), aka one-dimensional Convolutional Neural Networks (1D CNNs), are multilayer artificial neural network architectures to classify patterns with shift-invariance and model context at each layer of the network. ECAPA-TDNN is a new TDNN-based speaker-embedding extractor for speaker verification; it is built upon the original x-vector architecture and puts more emphasis on channel attention, propagation, and aggregation.
Implementation
After downloading the Common Voice dataset, the data is preprocessed by converting the MP3 files into WAV format to avoid information loss and separated into training, validation, and testing sets.
A pretrained VoxLingua107 model is retrained with the Common Voice dataset using the Hugging Face SpeechBrain library to focus on the languages of interest. VoxLingua107 is a speech dataset used for training spoken language recognition models that work well with real-world and varying speech data. This dataset contains data for 107 languages. By default, Japanese and Swedish are used, and more languages can be included. This model is then used for inference on the testing dataset or a user-specified dataset. Also, there is an option to utilize SpeechBrain’s Voice Activity Detection (VAD) where only the speech segments from the audio files are extracted and combined before samples are randomly selected as input into the model. This link provides all the necessary tools to perform VAD. To improve performance, the user may quantize the trained model to integer-8 (INT8) using Intel Neural Compressor to decrease latency.
Training
The copies of training scripts are added to the current working directory, including create_wds_shards.py – for creating the WebDataset shards, train.py – to perform the actual training procedure, and train_ecapa.yaml – to configure the training options. The script to create WebDataset shards and YAML file are patched to work with the two languages chosen for this code sample.
In the data preprocessing phase, prepareAllCommonVoice.py script is executed to randomly select a specified number of samples to convert the input from MP3 to WAV format. Here, 80% of these samples will be used for training, 10% for validation, and 10% for testing. At least 2000 samples are recommended as the number of input samples and is the default value.
In the next step, WebDataset shards are created from the training and validation datasets. This stores the audio files as tar files which allows writing purely sequential I/O pipelines for large-scale deep learning in order to achieve high I/O rates from local storage—about 3x-10x faster compared to random access.
The YAML file will be modified by the user. This includes setting the value for the largest number for the WebDataset shards, output neurons to the number of languages of interest, number of epochs to train over the entire dataset, and the batch size. The batch size should be decreased if the CPU or GPU runs out of memory while running the training script.
In this code sample, the training script will be executed with CPU. While running the script, “cpu” will be passed as an input parameter. The configurations defined in train_ecapa.yaml are also passed as parameters.
The command to run the script to train the model is:
python train.py train_ecapa.yaml --device "cpu"
In the future, the training script train.py will be designed to work for Intel® GPUs such as the Intel® Data Center GPU Flex Series, Intel® Data Center GPU Max Series, and Intel® Arc™ A-Series with updates from Intel Extension for PyTorch.
Run the training script to learn how to train the models and execute the training script. The 4th Generation Intel® Xeon® Scalable Processor is recommended for this transfer learning application because of its performance improvements through its Intel® Advanced Matrix Extensions (Intel® AMX) instruction set.
After training, checkpoint files are available. These files are used to load the model for inference.
Inference
The crucial step before running inference is to patch the SpeechBrain library’s pretrained interfaces.py file so that PyTorch TorchScript* can be run to improve the runtime. TorchScript requires the output of the model to be only tensors.
Users can choose to run inference using the testing set from Common Voice or their own custom data in WAV format. The following are the options the inference scripts (inference_custom.py and inference_commonVoice.py) can be run with:
| Input Option | Description |
| -p | Specify the data path. |
| -d | Specify the duration of wave sample. The default value is 3. |
| -s | Specify size of sample waves, default is 100. |
| –vad | (`inference_custom.py` only) Enable VAD model to detect active speech. The VAD option will identify speech segments in the audio file and construct a new .wav file containing only the speech segments. This improves the quality of speech data used as input into the language identification model. |
| –ipex | Run inference with optimizations from Intel Extension for PyTorch. This option will apply optimizations to the pretrained model. Using this option should result in performance improvements related to latency. |
| –ground_truth_compare | (`inference_custom.py` only) Enable comparison of prediction labels to ground truth values. |
| –verbose | Print additional debug information, like latency. |
The path to the data must be specified. By default, 100 audio samples of 3-seconds will be randomly selected from the original audio file and used as input to the language identification model.
A small Convolutional Recurrent Deep Neural Network (CRDNN) pretrained on the LibriParty dataset is used to process audio samples and output the segments where speech activity is detected. This can be used in inference with the --vad option.
From the figure below, the timestamps where speech will be detected is delivered from the CRDNN model, and these are used to construct a new, shorter audio file with only speech. Sampling from this new audio file will give a better prediction of the primary language spoken.
Run the inference script yourself. An example command of running inference:
python inference_custom.py -p data_custom -d 3 -s 50 --vad
This will run inference on data you provide located inside the data_custom folder. This command performs inference on 50 randomly selected 3-second audio samples with voice activity detection.
If you want to run the code sample for other languages, download Common Voice Corpus 11.0 datasets for other languages.
Optimizations with Intel Extension for PyTorch and Intel Neural Compressor
PyTorch
The Intel extension expands PyTorch with up-to-date features and optimizations for an extra performance boost on Intel hardware. Check out how to install Intel Extension for PyTorch. The extension can be loaded as a Python module or linked as a C++ library. Python users can enable it dynamically by importing intel_extension_for_pytorch.
- The CPU tutorial gives detailed information about Intel Extension for PyTorch for Intel CPUs. Source code is available at the master branch.
- The GPU tutorial gives detailed information about Intel Extension for PyTorch for Intel GPUs. Source code is available at the xpu-master branch.
To optimize the model for inference using Intel Extension for PyTorch, the --ipexoption can be passed in. The model is optimized using the plug-in. TorchScript speeds up inference because PyTorch is run in graph mode. The command to run with this optimization is:
python inference_custom.py -p data_custom -d 3 -s 50 --vad --ipex --verbose
Note: The --verbose option is required to view the latency measurements.
Auto-mixed precision such as bfloat16 (BF16) support will be added in a future release of the code sample.
Intel Neural Compressor
This is an open-source Python library that runs on CPUs or GPUs, which:
- Performs model quantization to reduce the model size and increase the speed of deep learning inference for deployment.
- Automates popular methods such as quantization, compression, pruning, and knowledge distillation across multiple deep-learning frameworks.
- Is part of the AI Kit
The model can be quantized from float32 (FP32) precision to integer-8 (INT8) by running the quantize_model.py script while passing in the path to the model and a validation dataset. The following code can be used to load this INT8 model for inference:
from neural_compressor.utils.pytorch import load
model_int8 = load("./lang_id_commonvoice_model_INT8", self.language_id)
signal = self.language_id.load_audio(data_path)
prediction = self.model_int8(signal)
Note that the original model is required when loading the quantized model. The command to quantize the trained model from FP32 to INT8 by using quantize_model.py is:
python quantize_model.py -p ./lang_id_commonvoice_model -datapath $COMMON_VOICE_PATH/commonVoiceData/commonVoice/dev
What’s Next?
Try out the above code sample by upgrading the hardware to a 4th Generation Intel Xeon Scalable Processor with Intel AMX and identify up to 93 different languages from Common Voice datasets.
We encourage you to learn more about and incorporate Intel’s other AI/ML Framework optimizations and end-to-end portfolio of tools into your AI workflow. Also, visit AI & ML page covering Intel’s AI software development resources for preparing, building, deploying, and scaling your AI solutions.
For more details about the new 4th Gen Intel Xeon Scalable processors, visit Intel’s AI Solution Platform portal where you can learn how Intel is empowering developers to run end-to-end AI pipelines on these powerful CPUs.
Useful resources
- Intel AI Developer Tools and resources
- oneAPI unified programming model
- Official documentation – Intel® Optimization for TensorFlow*
- Official documentation – Intel® Neural Compressor
- Accelerate AI Workloads with Intel® AMX
Explore more AI code samples
American Sign Language Fingerspelling Recognition
Posted by Thad Starner (Professor, Georgia Tech and Staff Research Scientist, Google), Sam Sepah (ML Research Program Manager), Manfred Georg (Software Engineer, Google), Mark Sherwood (Senior Product Manager, Google), Glenn Cameron (Product Marketing Manager, Google)
Over 70 million deaf people around the world use sign language to communicate. Collectively, they use more than 300 different sign languages worldwide. And over 1.5 billion people are affected by hearing loss globally. Most Deaf and Hard of Hearing people cannot use their voice to initiate a search or perform actions due to speech limitations. Additionally, the interfaces used by smart home devices and mobile platforms to respond to speech are generally audio based.
Signed languages are sophisticated systems of communication, each with a complete set of language features. On a surface level, handshapes along with four other “parameters” form the basis of signed communication. An open hand or a closed hand while making the same motion can completely change the meaning of a sign. Likewise, palm orientation, motion/contact, location, and non-manual markers (typically mouth movements and facial expressions) define individual signs. A number of grammatical constructs, some of which have no analog in spoken languages, allow a signer to produce complex phrases.
As we develop translation systems for American Sign Language (ASL) and other sign languages, it is natural to break apart various aspects of the language and attempt to perform tasks using those parts.
To that end, we’re excited to announce the release of one of the largest datasets of ASL fingerspelling and a Kaggle ML competition that will award $200k in prizes to ML engineers who develop the most accurate ASL fingerspelling recognition models using MediaPipe and TensorFlow Lite. The winning models will be open sourced to help developers add support for fingerspelling to their apps.
| Watch These Hands (Kaggle remix) Performed by Sean Forbes, Co-Founder, Deaf Professional Arts Network |
Fingerspelling communicates words using hand shapes that represent individual letters. While fingerspelling is only a part of sign languages, it is often used for communicating names, addresses, phone numbers, names, and other information that is commonly entered on a mobile phone. Many Deaf smartphone users can fingerspell words faster than they can type on mobile keyboards. In fact, in our dataset, ASL fingerspelling of phrases averages 57 words per minute, which is substantially faster than the US average of 36 words per minute for an on screen keyboard. But, sign language recognition AI for text entry lags far behind voice-to-text or even gesture-based typing, as robust datasets didn’t previously exist.
Although fingerspelling is just a small part of sign languages, there are many reasons to produce systems which specifically focus on it, even while maintaining an ultimate goal of full translation. While fingerspelling at full speed (which can peak over 80 words per minute) the handshapes in the fingerspelling co-articulate together and entire words can become lexicalized into different shapes from their slowed down version. The resulting movements are visually among the fastest used in ASL, and thus stretch particular aspects of any visual recognition system which seeks to perform full translation.
Big Steps Forward
Google Research and the Deaf Professional Arts Network have worked together to create a massive fingerspelling dataset that we will release for this competition to help move sign language recognition forward. The dataset includes over 3 million fingerspelled characters produced by over 100 Deaf signers in the form of continuous phrases, names, addresses, phone numbers, and URLs. This signing was captured using the selfie camera of a smartphone with a variety of backgrounds and lighting conditions and is the largest dataset collection of its kind to date.
Large language models show increasing promise in a variety of language and speech tasks. Everything from chat agents to assistant technology is progressing at breathtaking speed. It is time to ensure that gesture and visual based systems also produce usable interfaces. Fingerspelling recognition models are part of this larger solution, which will address the widening gap in accessibility for Deaf and Hard of Hearing individuals.
How to Get Involved
Join the Kaggle competition today to help us make AI more accessible for the Deaf and hard of hearing community.
Differential privacy for deep learning at GPT scale
Two papers from Amazon Web Services AI present algorithms that alleviate the intensive hyperparameter search and fine-tuning required by privacy-preserving deep learning at very large scales.Read More
Differential privacy for deep learning at GPT scale
Two new methods — automatic gradient clipping and bias-term-only fine-tuning — improve the efficiency of differentially private model training.Read More
Startup’s AI Slashes Paperwork for Doctors Across Africa
As a medical doctor in Nigeria, Tobi Olatunji knows the stress of practicing in Africa’s busy hospitals. As a machine-learning scientist, he has a prescription for it.
“I worked at one of West Africa’s largest hospitals, where I would routinely see more than 30 patients a day — it’s a very hard job,” said Olatunji.
The need to write detailed patient notes and fill out forms makes it even harder. Paper records slowed the pace of medical research, too.
In his first years of practice, Olatunji imagined a program to plow through the mounds of paperwork, freeing doctors to help more patients.
It’s been a journey, but that software is available today from his company, Intron Health, a member of the NVIDIA Inception program, which nurtures cutting-edge startups.
A Side Trip in Tech
With encouragement from med school mentors, Olatunji got a master’s degree in medical informatics from the University of San Francisco and another in computer science at Georgia Tech. He started working as a machine-learning scientist in the U.S. by day and writing code on nights and weekends to help digitize Africa’s hospitals.
A pilot test during the pandemic hit a snag.
The first few doctors to use the code took 45 minutes to finish their patient notes. Feeling awkward in front of a keyboard, some health workers said they prefer pen and paper.
“We made a hard decision to invest in natural language processing and speech recognition,” he said. It’s technology he was already familiar with in his day job.
Building AI Models
“The combination of medical terminology and thick African accents produced horrible results with most existing speech-to-text software, so we knew there would be no shortcut to training our own models,” he said.
The Intron team evaluated several commercial and open-source speech recognition frameworks and large language models before choosing to build with NVIDIA NeMo, a software framework for text-based generative AI. In addition, the resulting models were trained on NVIDIA GPUs in the cloud.
“We initially tried to train with CPUs as the cheapest option, but it took forever, so we started with a single GPU and eventually grew to using several of them in the cloud,” he said.
The resulting Transcribe app captures doctors’ dictated messages with more than 92% accuracy across more than 200 African accents. It slashes the time they spend on paperwork by 6x on average, according to an ongoing study Intron is conducting across hospitals in four African countries.
“Even the doctor with the fastest typing skills in the study got a 40% speedup,” he said of the software now in use at several hospitals across Africa.
Listening to Africa’s Voices
Olatunji knew his models needed high quality audio data. So, the company created an app to capture sound bites of medical terms spoken in different accents.
To date, the app’s gathered more than a million clips from more than 7,000 people across 24 countries, including 13 African nations. It’s one of the largest datasets of its type, parts of which have been released as open source to support African speech research.
Today, Intron refreshes its models every other month as more data comes in.
Nurturing Diversity in Medtech
Very little research exists on speech recognition for African accents in a clinical setting. So, working with Africa’s tech communities like DSN, Masakhane and Zindi, Intron launched AfriSpeech-200, a developer challenge to kickstart research using its data.
Similarly, for all its sophistication, medtech lags in diversity and inclusion, so Olatunji recently launched an effort that addresses that issue, too.
Bio-RAMP Lab is a global community of minority researchers working on problems they care about at the intersection of AI and healthcare. The group already has a half dozen papers under review at major conferences.
“For seven years, I was the only Black person on every team I worked on,” he said. “There were no Black scientists or managers, even in my job interviews.”
Meanwhile, Intron is even helping hospitals in Africa find creative ways to acquire the hardware they need. It’s another challenge on the way to opening up huge opportunities.
“Once healthcare data gets digitized, you unlock a whole new world for research into areas like predictive models that can be early warning systems for epidemics — we can’t do it without data,” Olatunji said.
Watch a masterclass (starting at 20:30) with Olatunji, HuggingFace and NVIDIA on AI for speech recognition.
100 things we announced at I/O 2023
Google I/O 2023 was filled with news and launches — here are 100 things announced at I/O.Read More
Time to Prioritize: Upgrade to Priority at 40% Off This GFN Thursday
Make gaming a priority this GFN Thursday — time’s running out to upgrade to a GeForce NOW Priority six-month membership at 40% off the normal price. Find out how new Priority members are using the cloud to get their game on.
Plus, the week brings updates for some of the hottest games in the GeForce NOW library, and four more titles join the list.
GeForce NOW RTX 4080 SuperPODs are now live for Ultimate members in Atlanta, where the gamers game. Follow along with the server rollout, and upgrade today for the Ultimate cloud gaming experience.
Priority Check
Through Sunday, May 21, save 40% on a six-month Priority membership for $29.99, normally $49.99.
Priority memberships are perfect for those looking to try GeForce NOW or lock in a lower price for a half-year. Priority members get higher access to GeForce gaming servers, meaning less wait times than free members.
Members who claimed this offer in its first week alone played over 1,000 different titles in the GeForce NOW library, for 30,000+ streamed hours. That means these Priority members skipped the line by more than 500 hours.
They also played the best of PC gaming across multiple devices — PCs, Macs, mobile devices and smart TVs, plus new categories of devices made possible by the cloud, like gaming Chromebooks and cloud gaming handheld devices. And they experienced the cinematic quality of RTX ON in supported titles.
With more than 1,600 titles in the GeForce NOW library, there’s something for everyone to play. Jump into squad-based action in Fortnite or Destiny 2, bring home the victory League of Legends or Counter-Strike: Global Offensive, and explore in open-world role-playing games like Genshin Impact and Cyberpunk 2077. With GeForce NOW Priority, members can get straight into the action.
But don’t wait: This offer ends on Sunday, May 21, so make it a priority to upgrade today.
Game On
GFN Thursday means more games for more gamers. This week brings new additions to the GeForce NOW library, and new updates for the hottest games.
Apex Legends: Arsenal, the latest season in EA and Respawn Entertainment’s battle royale FPS, is available this week for GeForce NOW members. Meet the newest playable Legend, Ballistic, who’s come out of retirement to teach the young pups some respect. Battle through an updated World’s Edge map, hone your skills in the newly updated Firing Range and progress through the new Weapon Mastery system.
In addition, Occupy Mars, the latest open-world sandbox game from Pyramid Games, joins the GeForce NOW library this week. Explore and colonize Mars, building a home base and discovering new regions. Grow crops, conduct mining operations and survive on an unforgiving planet. As all sci-fi films that take place on Mars have shown, things don’t always go as planned. Players must learn to cope and survive on the red planet.
For more action, take a look at what’s joining the GeForce NOW library this week:
- Voidtrain (New release on Steam, May 9)
- Occupy Mars: The Game (New release on Steam, May 10)
- Far Cry 6 (New Release Steam, May 11)
- TT Isle of Man: Ride on the Edge 3 (New release on Steam, May 11)
Ultimate members can now enable real-time ray tracing in Fortnite. The island’s never looked so good.
What are you playing this weekend? We’ve got a little challenge for you this week. Let us know your response on Twitter or in the comments below.
Replace a word in a video game title with “cloud.”
We’ll go first: Cloud of Legends
—
NVIDIA GeForce NOW (@NVIDIAGFN) May 10, 2023
Living on the Edge: Singtel, Microsoft and NVIDIA Dial Up AI Over 5G
For telcos around the world, one of the biggest challenges to upgrading networks has always been the question, “If you build it, will they come?”
Asia’s leading telco, Singtel, believes the key to helping customers innovate with AI across industries — for everything from traffic and video analytics to conversational AI avatars powered by large language models (LLMs) — is to offer multi-access edge compute services on its high-speed, ultra-low-latency 5G network.
Multi-access edge computing, or MEC, moves the computing of traffic and services from a centralized cloud to the edge of the network, where it’s closer to the customer. Doing so reduces network latency and lowers costs through sharing of network resources.
Singtel is collaborating with Microsoft and NVIDIA to combine AI and 5G, so enterprises can boost their innovation and productivity. Using NVIDIA’s full-stack accelerated computing platform optimized for Microsoft Azure Public MEC, the telco is creating solutions that enable customers to leverage AI video analytics for multiple use cases and to deploy 5G conversational avatars powered by LLMs.
From Sea to Shore
Singtel has been rolling out enterprise 5G and MEC across ports, airports, manufacturing facilities and other locations. In addition to running low-latency applications at the edge using Singtel’s 5G network, the solution has the potential to transform operations in sectors such as public safety, urban planning, healthcare, banking, civil service, transportation and logistics. It also offers high security for public sector customers and better performance for end users, enabling new intelligent edge scenarios.
Customers can use these capabilities through Microsoft Azure, only paying for the amount of compute and storage they use for the duration in which they use it. This replicates the cloud consumption model at the network edge and lets users save on additional operational overhead.
Edge Technologies
Singtel is working with video analytics software-makers participating in NVIDIA Inception, a free program that offers startups go-to-market support, expertise and technology. These ISVs will be able to use the NVIDIA Jetson Orin module for edge AI and robotics in conjunction with Microsoft MEC to identify traffic flows at airports and other high-population areas, retail video analytics and other use cases.
Singtel and NVIDIA are also showcasing their technology and solutions, including a real-time LLM-powered avatar developed by system integrator Quantiphi and based on NVIDIA Omniverse digital twin technology, at a May 11 launch event in Singapore. The avatar, built with NVIDIA Riva speech AI and the NeMo Megatron transformer model, enables people to interact in natural language on any topic of interest. Businesses can deploy these avatars anywhere over 5G.
Using Singtel’s high-speed, low-latency 5G — combined with NVIDIA AI accelerated infrastructure and capabilities — enterprises can explore use cases on everything from computer vision and mixed reality to autonomous guided vehicles.
Singtel plans to expand these new capabilities beyond Singapore to other countries and affiliated telcos, as well. This collaboration will help redefine what’s possible through the powerful combination of compute and next-generation networks, unlocking new operational efficiencies, revenue streams and customer experiences.