Intel's oneDNN 2.1 Released With NVIDIA GPU Support, Initial Alder Lake Optimizations

Out today is a new release of Intel's open-source oneDNN library used as a deep neural network library for assembling deep learning applications. With the new oneDNN 2.1 release there is now initial support for NVIDIA GPU acceleration as well as a host of improvements for running on forthcoming Intel CPUs.

oneDNN 2.1 is the latest feature-packed update to this oneAPI component formerly known as DNNL and before that was MKL-DNN. The oneDNN 2.1 release has a variety of improvements when running on Intel Gen / Xe Graphics, mostly various performance optimizations. On the Intel CPU front is also a number of optimizations including BFloat16 for Intel Xeon Scalable CPUs supporting the Advanced Matrix Extensions (AMX, coming with Sapphire Rapids), CPU ISA hints support, various AVX-512 optimizations, and initial support for processors with AVX2 and DL-BOOST (the combination debuting with Alder Lake). There is even some INT8 optimizations for CPUs with SSE4.1 and a number of other low-level improvements.

With oneAPI/oneDNN being open-source and being used on more than just Intel hardware, there are even more AArch64 (ARM 64-bit) enhancements in the 2.1 release. The oneDNN 2.1 release has various performance improvements with ArmCL. There is also now JIT support for AArch64 along with implementations for various primitives.

A new preview-level feature with oneDNN 2.1 is support for NVIDIA GPUs. The oneDNN library now supports NVIDIA GPU acceleration when using the proprietary driver stack with the cuDNN and cuBLAS libraries. Targeting the NVIDIA GPUs relies on using Intel's DPC++ Compiler.

Download links and all the details on the oneDNN 2.1 release can be found via GitHub.

For those wondering about oneDNN performance on various Intel and non-Intel devices, see the oneDNN reference benchmarks on OpenBenchmarking.org for various oneDNN primitives.