oneDNN v3.14.1 patch addresses Intel GPU and x64 CPU matmul correctness

The UXL Foundation has released oneDNN version 3.14.1, a patch update focused on improving the mathematical accuracy and performance of deep learning primitives. This release specifically targets correctness bugs in matrix multiplication (matmul) operations occurring on x64 CPUs and Intel GPUs.

Users running AI workloads on Intel hardware will benefit from fixes involving binary post-operations and sparse groups. Additionally, the update includes performance optimizations for matrix-vector multiplication on x64 architectures and updates the minimum supported version for ZenDNN to v6.1.2.

Technical improvements in oneDNN v3.14.1

• Intel GPU Optimizations: The release fixes correctness issues in grouped matmul involving sparse groups and post-operations, while significantly improving primitive creation performance for grouped matmul cases.
• x64 CPU Correctness: Multiple fixes address matmul accuracy, specifically regarding binary post-ops and batched f32 matmul operations utilizing the Intel AVX-512 instruction set.
• Performance Tuning: Improvements were implemented for f32 matmul performance during matrix-vector multiplication scenarios on x64 architectures.
• Dependency and Tooling Updates: The minimal supported version for ZenDNN has been bumped to v6.1.2, and a memory allocation bug in the benchdnn tool for small tensors in cold-cache mode has been resolved.

Why it matters

As a core component of the oneAPI ecosystem, these fixes ensure more reliable AI inference and training results for developers using Intel CPUs and GPUs. By resolving AVX-512 and sparse group inaccuracies, oneDNN maintains its role as a stable, high-performance backend for open-source AI frameworks targeting heterogeneous hardware.