Mahalagang impormasyon
Intel and AMD are collaborating to integrate their CPU optimizations for PyTorch into a single, unified library. AMD is transitioning its ZenDNN library into the open-source oneDNN project, which is managed by the UXL Foundation. Additionally, AMD has joined the UXL Foundation as a steering member with an annual contribution of $20,000.
This initiative, recommended by the x86 Ecosystem Advisory Group established in 2024, aims to simplify AI inference for developers. By utilizing the oneDNN backend, software can automatically detect hardware features at runtime to generate optimized kernels. The library supports essential machine learning operations, including matrix multiplication and convolution, while providing low-precision data types.
The oneDNN backend, which was integrated into PyTorch in late 2023, has demonstrated significant performance gains. According to reported data, the library achieved a 1.24x improvement for Float32 data types and a 2.21x increase for BFloat16 data types on supported hardware.
While this effort focuses on x86 architectures, oneDNN remains a cross-platform library that also supports Arm and RISC-V. Intel, a founding member of the UXL Foundation alongside companies like Google Cloud and Qualcomm, originally developed oneDNN before transitioning it to the foundation.
