AMD and Intel have announced an important collaboration to unify oneDNN in PyTorch and optimize their software optimizations for artificial intelligence on x86 processors. The move centralizes the work in the open-source oneDNN library, governed by the Unified Acceleration Foundation (UXL), making AI application development simpler and more consistent across the two architectures.
What happened?
The two companies, historic rivals in the CPU market, have aligned on oneDNN as the default open-source library for x86 acceleration in PyTorch. AMD will integrate the optimizations from its ZenDNN directly into oneDNN and has joined the UXL Foundation as a steering member to advance open collaboration. This union eliminates the need to maintain separate code paths for each manufacturer.
The initiative was recommended by the x86 Ecosystem Advisory Group and aims to reduce fragmentation in the AI ecosystem. Developers now have a common software path to optimize AI workloads on AMD and Intel processors, facilitating portability and project maintenance.
Why does it matter?
Maintaining separate code paths for each vendor creates cost, complexity, and additional development time. With the unification of oneDNN in PyTorch, applications based on this library become easier to build, maintain, and deploy on any modern x86 CPU. This is especially relevant for inference and training workloads that run locally or on mixed servers.
The companies are also collaborating on AI Compute Extensions (ACE), a new x86 instruction set extension designed to accelerate matrix multiplication, a core operation in AI workloads, with future support via oneDNN updates. Robert Hormuth, corporate vice president of Architecture and Strategy at AMD, said the company is advancing an open software ecosystem that gives developers greater choice. Jeff McVeigh of Intel emphasized that developers should not have to think about separate software paths to get the best from an x86 CPU.
Context and practical impact
oneDNN, originally developed by Intel, is already widely used by frameworks such as PyTorch, TensorFlow, and ONNX Runtime for deep learning primitives on CPUs. AMD maintained a compatible fork called ZenDNN. The upstream integration eliminates this duplication and accelerates the adoption of optimizations specific to each microarchitecture.
In practice, this means a model trained or optimized on an AMD system can run competitively on Intel, and vice versa, without extensive manual adjustments. The collaboration strengthens x86's position against architectures like Arm in AI workloads, especially in data centers and edge computing.
Sources: AMD Newsroom, Tom's Hardware, and industry reports.
By GeekikiBot
