Intel and AMD to merge x86 CPU optimizations into oneDNN for PyTorch
AMD will upstream its ZenDNN optimizations into Intel's oneDNN library and join the UXL Foundation as a steering member, ending its separate fork for PyTorch x86 acceleration.
Summary
Intel and AMD announced joint collaboration to consolidate x86 CPU optimization work for PyTorch into oneDNN, the open-source neural network library Intel originally developed. AMD will contribute its ZenDNN improvements directly upstream rather than maintaining a separate API-compatible fork. AMD is also joining the UXL Foundation, the consortium that oversees oneDNN and other former oneAPI projects, as a steering member.
What changes in the oneDNN collaboration
- oneDNN becomes the unified default for x86 acceleration in PyTorch, replacing AMD's separate ZenDNN path
- ZenDNN optimizations upstreamed: AMD's Zen-specific performance work will be merged into oneDNN rather than maintained as a downstream fork
- UXL Foundation expansion: AMD joins as steering member alongside Intel, Samsung, Qualcomm, Arm and others
- Follows x86 Ecosystem Advisory Group: the collaboration extends prior ISA-level cooperation between the rivals
Why it matters
A unified oneDNN path reduces fragmentation for PyTorch users on x86 hardware and may accelerate optimization delivery for both Intel and AMD processors. AMD's UXL Foundation membership signals deeper commitment to the open governance of Intel-originated AI infrastructure.
What to do
- Wait for AMD's ZenDNN optimizations to be upstreamed into oneDNN if you use PyTorch on AMD Zen CPUs.
- Expect oneDNN to become the default x86 acceleration library in PyTorch, replacing the separate ZenDNN path.
- Watch for performance and compatibility improvements in PyTorch x86 CPU workloads as Intel and AMD consolidate their optimization efforts.
Original reference: www.phoronix.com →