THE CRUNCH
AMD and Intel are pooling their CPU optimisations for PyTorch into a single library. AMD has maintained ZenDNN, its own downstream version of Intel's oneDNN, and that work is now moving into the open-source oneDNN project, which is published by the UXL Foundation and driven by Intel. The result is one code path that tunes AI inference for both companies' chips.
The unification was recommended by the x86 Ecosystem Advisory Group, a joint initiative primarily between Intel and AMD, established in 2024. The two rivals have collaborated through it before, notably on the ACE extensions for x86. AMD also announced it is joining the UXL Foundation as a steering member, a role it says influences the foundation's future in exchange for $20,000 per year.
oneDNN serves as the backend for CPU inference in frameworks like PyTorch, detecting hardware features at runtime to generate optimised kernels and supporting low-precision data types. When it was added to PyTorch in late 2023, it showed a 1.24x improvement for Float32 workloads and 2.21x for BFloat16. For developers, the practical upshot is a single backend that applies the right optimisations whatever x86 chip the code runs on.
Although the focus here is x86, oneDNN is a cross-platform library that also supports Arm and RISC-V architectures.