AMD and Intel are unifying their software efforts to streamline AI inference on x86 processors. The companies will integrate AMD’s ZenDNN library into Intel’s oneDNN, creating a single code path for PyTorch. This initiative was recommended by the x86 Ecosystem Advisory Group, a joint venture established by the two chipmakers in 2024.
The oneDNN backend is designed to accelerate CPU inference by detecting hardware features at runtime and generating optimized kernels. It supports common machine learning operations, including matrix multiplication and convolution, while providing low-precision data types. The library is cross-platform and also supports Arm and RISC-V architectures.
When oneDNN was initially integrated into PyTorch in late 2023, it demonstrated performance gains of 1.24x for Float32 data types and 2.21x for BFloat16. By consolidating these efforts, developers can now rely on a single backend to deliver hardware-specific optimizations across different systems.
Intel originally developed oneDNN before contributing it to the UXL Foundation. The foundation includes several major technology companies as founding members, such as Arm, Qualcomm, and Google Cloud, aimed at fostering open-standard software development.
