AMD has gained progress in building a robust software stack that supports an open ecosystem of models, libraries, frameworks, and tools. With proven platforms gaining momentum, there is significance of a leadership software stack and an optimized ecosystem for achieving application performance. PyTorch is a key part of AMD’s AI journey, and AMD’s Victor Peng, AMD President and Soumith Chintala, founder of PyTorch discussed the latest progress at the DC & AI Keynote on June 12.
Building a Powerful SW Stack with ROCm
Victor introduced ROCm, AMD’s SW stack for Instinct Data Center GPUs. It offers a comprehensive set of open-source libraries, runtime, compilers, and tools for developing, running, and fine-tuning AI models. The fifth generation ROCm incorporates optimizations for AI and high-performance computing workloads, including tailored kernels for low-latency memory systems, support for new data types, and integration with OpenAI Triton. With tools for porting AI software to AMD Instinct platforms, ROCm ensures quality and robustness, tested extensively and compliant with PyTorch and TensorFlow frameworks.
Collaboration with PyTorch
To shed light on the partnership between AMD and PyTorch, Victor invited Soumith Chintala, the founder of PyTorch, to discuss the advancements and integration between the two. PyTorch, the industry’s most famous AI framework, boasts a vibrant developer community and is known for its continuous innovation and incorporation of cutting-edge research.
To highlight the AMD and PyTorch partnership, Victor hosted a discussion with Soumith Chintala, the founder of PyTorch. PyTorch, renowned for its innovation and community, is the industry’s leading AI framework. The latest version, PyTorch 2.0, integrates with hardware-agnostic software compilers like OpenAI Triton, enabling efficient training and deployment of AI models. With optimized techniques, PyTorch 2.0 enhances productivity and offers remarkable speed improvements. The collaboration between AMD and the PyTorch Foundation ensures seamless utilization of AMD GPUs, expanding AI accelerator accessibility worldwide and paving the way for future optimizations and broader hardware support.
Empowering the Developer Community
The partnership between AMD and PyTorch benefits the developer community by democratizing access to AI accelerators. Support for AMD GPUs in PyTorch allows developers to train and deploy models across various platforms, including CPUs like EPYC and Ryzen, GPUs like Instinct and Radeon, and embedded devices like Versal SoCs. By ensuring immediate compatibility of new models on AMD platforms, the collaboration streamlines the development process and empowers developers to leverage the full potential of AMD’s hardware. This increased accessibility and flexibility enable developers worldwide to push the boundaries of AI innovation.
Hugging Face and AI Model Innovation
Victor praised Hugging Face as the leading force behind open-source AI model innovation, empowering generative AI with transformative transformers. AMD’s optimized software enables a high-performing development stack, supporting groundbreaking AI advancements for customers and developers through scalable real-world deployments.
Conclusion
At the DC & AI Keynote, AMD demonstrated its dedication to openness, performance, and collaboration. The ROCm SW stack, PyTorch integration, and support for Hugging Face exemplify AMD’s commitment to empowering developers and researchers to achieve AI breakthroughs. By offering accessible, high-performing solutions, AMD fuels the future of AI as a leading GPU platform integrated with PyTorch.
To listen to the full keynote visit the AMD Youtube channel
To listen to Soumith Chintala’s section of the keynote