Hugging Face and AMD today announced a strategic partnership to deliver state-of-the-art transformer performance on AMD CPUs and GPUs. The collaboration, unveiled at AMD's Data Center and AI Technology Premiere, aims to provide developers and organizations with more hardware options for training and inference, potentially lowering costs and improving performance.
Under the partnership, Hugging Face will optimize its open-source libraries for AMD hardware, starting with the ROCm SDK integration. Initial tests show AMD's MI250 GPU trains BERT-Large 1.2x faster and GPT2-Large 1.4x faster than competing hardware. The collaboration will support enterprise-grade Instinct MI2xx and MI3xx GPUs, consumer Radeon Navi3x GPUs, Ryzen and EPYC CPUs, and Alveo V70 AI accelerators.
Supported model architectures include BERT, GPT2, T5, LLaMA, Vision Transformer, and many others across natural language processing, computer vision, and speech. The companies will also validate frameworks like PyTorch, TensorFlow, and ONNX Runtime.
The partnership aims to create a dedicated Optimum library for AMD platforms, enabling seamless use with minimal code changes. "Open-source means the freedom to build from a wide range of software and hardware solutions," said Hugging Face CEO Clement Delangue.
This collaboration addresses growing concerns over limited deep learning hardware options and rising costs, with the goal of setting new cost-performance standards.