Released on Friday was a new version of the oneDNN deep neural network library maintained by Intel and the UXL Foundation. This library used by various deep learning applications continues preparing for upcoming Intel CPUs and GPUs.
The oneDNN 3.9 release candidate is now available for testing. The 3.9 series is bringing initial support for next-generation Intel Xeon Diamond Rapids processors with Intel AVX 10.2 and Intel AMX support. This Diamond Rapids “DMR” support isn’t yet enabled by default but requires use of the “ONEDNN_MAX_CPU_ISA=AVX10_2_512_AMX_2” option.
The oneDNN 3.9 release on the client side also brings initial support for Nova Lake processors with AVX 10.2 instruction set support. That Nova Lake support is currently gated by the “ONEDNN_MAX_CPU_ISA=AVX10_2_512” option.
Over on the graphics side, the oneDNN library has been working out performance improvements for upcoming Xe3 graphics hardware. There are also performance improvements for existing Lunar Lake SoCs around MATMUL performance.
Outside of the Intel space, oneDNN 3.9 continues optimizing its AArch64 (ARM64) performance.
The oneDNN 3.9 release is also working on FP8 performance improvements, better MATMUL primitive performance for CPUs with Intel AMX, and other optimizations.
More details on these oneDNN 3.9-rc changes via GitHub.
