AMD and Intel have merged their CPU optimizations into the oneDNN library, creating a single PyTorch backend that boosts inference performance on both AMD and Intel CPUs.
Unified Backend for PyTorch
- oneDNN now supports both AMD and Intel x86 CPUs via a single code path.
- AMD’s ZenDNN branch has been incorporated into the open‑source oneDNN repository.
- Both companies contribute to the UXL Foundation, with AMD paying $20,000 annually.
Key Technical Highlights
| Aspect | Details |
|---|---|
| Performance Gains | 1.24× Float32, 2.21× BFloat16 on latest CPUs |
| Runtime Detection | Hardware features auto‑detected for optimized kernels |
| Cross‑Platform | Works on x86, Arm, RISC‑V |
| Foundation | Founded by Intel, Arm, Qualcomm, Google Cloud, etc. |
Market Implications
The unified oneDNN backend simplifies developer workflows, reduces fragmentation, and positions AMD and Intel as collaborative leaders in AI inference on CPUs, potentially accelerating adoption of AI workloads in data centers and edge devices.