AMD and Intel Join Forces to Unify PyTorch Optimization Path on x86
AMD and Intel announced a collaboration on PyTorch x86 CPU optimization, integrating their respective optimizations into the open-source deep learning library oneDNN, making it the default open-source library for x86 acceleration within PyTorch. AMD will integrate ZenDNN optimizations into oneDNN and join the Unified Acceleration Foundation (UXL Foundation) as a steering committee member. The two companies are also advancing the next-generation instruction set architecture for x86, "AI Compute Extensions" (ACE), focusing on accelerating matrix multiplication, with plans to integrate it into PyTorch through oneDNN updates in the future.
Body · AI translation
On October 21, AMD and Intel announced yesterday (October 20) that they are collaborating on PyTorch x86 CPU optimization: the two companies will integrate their respective optimization efforts into the open-source deep learning library oneDNN, providing developers with a unified software path across both companies' CPUs.
Note: PyTorch is an open-source Python deep learning framework, mainly used for building, training, and deploying neural networks. It is known for its flexibility, ease of debugging, and Pythonic programming style, and is widely used in academic research and commercial AI applications.
According to the announcement, the two companies plan to concentrate their respective processor optimization efforts into the open-source library oneDNN, making it the default open-source library for x86 acceleration within PyTorch. AMD will integrate its ZenDNN optimizations into oneDNN and join the Unified Acceleration Foundation (UXL Foundation) as a steering committee member.
Driven by the x86 Ecosystem Advisory Group, this rare collaboration between the two companies was facilitated, "turning competition into collaboration" in the CPU software ecosystem. If implemented smoothly, developers, cloud service providers, and enterprise users may all gain a more consistent PyTorch performance experience on x86 servers, workstations, and AI PCs.
In addition to oneDNN, AMD and Intel are also advancing "AI Compute Extensions" (ACE). This is a next-generation instruction set architecture for x86, focused on accelerating matrix multiplication (a core operation in AI model training and inference), and related capabilities are planned to be integrated into PyTorch through future oneDNN updates.
For developers, PyTorch workloads that previously may have required separate adaptation and tuning for AMD or Intel CPUs are expected in the future to achieve optimized support through the same software path. The two companies say this will reduce fragmentation in the x86 AI ecosystem, lower duplicate development and maintenance costs, and make building, deploying, and migrating AI applications across platforms smoother.