Intel oneDNN 2.7 Released With Sapphire Rapids & DG2 Optimizations, AMD GPU Bringup

Aligned with Intel's Innovation event happening this week in San Jose, Intel on Tuesday released oneDNN 2.7 as the newest version of their deep neural network library. In addition to optimizing support for new Intel hardware, oneDNN 2.7 also has AMD GPU support.

Intel's oneAPI Deep Neural Network Library provides various fundamentals for deep learning / AI applications as a framework for helping to accelerate AI workloads both with CPUs and GPUs/accelerators. While Intel-focused, prior releases have brought support for AArch64 and RISC-V CPUs, OpenPOWER ISA support, IBM z/390x, NVIDIA GPUs, and now with oneDNN 2.7 is also initial AMD GPU support. The oneDNN 2.7 release has initial support for AMD GPUs by leveraging the AMD GPUOpen MIOpen library and supported primitives currently include local response normalizatiom, softmax, and eltwise.

The oneDNN 2.7 release also has new optimizations for Intel Xeon Scalable "Sapphire Rapids", including more preparations around AMX support. On the graphics side, there are performance improvements for Xe Architecture graphics, including Ponte Vecchio. There is also improvements benefiting Intel Arc DG2/Alchemist hardware too. Beyond the Intel optimizations and initial AMD GPU support, there are also a number of AArch64 CPU optimizations in this release too.

Those wanting to learn more about oneDNN 2.7 can do so via GitHub. I'll be upgrading to v2.7 soon for my oneDNN benchmarks.