oneDNN v3.14.2 improves FP8 support for Intel AVX-512 and fixes Intel GPU errors
This patch release enhances 8-bit floating point (FP8) stability and performance for Intel AVX-512 processors and resolves correctness issues on Intel GPUs.
Summary
The UXL Foundation has released version 3.14.2 of oneDNN, a patch update to the previous 3.14.1 release. This update focuses on improving the reliability and functionality of 8-bit floating point (FP8) operations, which are critical for modern AI workloads and efficient inference.
The release specifically targets stability for Intel AVX-512 enabled processors and Intel GPUs. It addresses several runtime crashes and correctness issues related to matrix multiplication, convolutions, and Scaled Dot Product Attention (SDPA) subgraphs.
Technical improvements in oneDNN v3.14.2
- FP8 Optimization: Added support for f8 destination in layer normalization forward propagation specifically for processors utilizing the Intel AVX-512 instruction set.
- Stability Fixes: Resolved runtime errors and system crashes occurring during matmul and convolution operations when using f8_e4m3 or f8_e5m2 binary post-operations on AVX-512 hardware.
- GPU Correctness: Fixed a specific correctness issue within the SDPA (Scaled Dot Product Attention) subgraph when handling a query length of 16 on Intel GPUs.
- Test Suite Cleanup: Removed benchdnn test cases involving mish post-ops and dynamic scales to prevent false positive correctness reports.
Why it matters
By stabilizing FP8 formats like e4m3 and e5m2, this update enables more reliable low-precision AI inference on Intel CPU architectures. The fixes for Intel GPUs ensure that developers building AI models using Scaled Dot Product Attention encounter fewer mathematical errors during execution.
Original reference: github.com →