Released one Friday was the newest version of oneDNN as this library started off by Intel and now officially under the UXL Foundation umbrella for serving as building blocks for deep learning software.
While under the UXL Foundation umbrella and supporting other hardware from Arm CPUs to NVIDIA GPUs, oneDNN 3.10 is overwhelmingly still about optimizing for Intel hardware. With oneDNN 3.10 a big focus has remained on preparing for future Intel CPUs with AVX 10.2 as well as better enhancing the Advanced Matrix Extensions (AMX) performance.
The oneDNN 3.10 release has improved performance for Intel Xeon CPUs with AVX 10.2 and AMX. On the Intel Core processor side is working on improved performance for future Core CPUs with AVX 10.2 support. The release notes do not detail the hardware generation for Intel Core CPUs with AVX 10.2 but simply cites using the “ONEDNN_MAX_CPU_ISA=AVX10_2_512”. AVX 10.2 was anticipated to appear with Nova Lake processors but the recent Nova Lake enablement in the GCC compiler went without any AVX10 mentions at this point.
The oneDNN 3.10 library also has improved MATMUL performance for CPUs with Intel AMX and a variety of other performance optimizations are in tow for Intel CPUs. On the Intel graphics side, oneDNN 3.10 has improved GEMM performance for small batch sizes on Lunar Lake, better MATMUL performance for Qwen2-7B shapes on Intel Arc Graphics, and other improvements.
For AArch64 processors are also a few performance optimizations too. Downloads and more details on the oneDNN 3.10 release via GitHub.
