Intel Releases oneDNN 2.0 To Bring The Open-Source Neural Network Library To Its GPUs

Intel's Deep Neural Network Library currently known as oneDNN as part of the oneAPI suite (and formerly known as MKL-DNN and DNNL) has reached version 2.0 as an open-source project.

This neural network library has long provided the "building blocks" for deep learning applications with very fast performance across x86_64 processors. The oneDNN library performs very well with these neural network primitives and seems to be gaining a fair amount of industry traction. With the continued adoption, oneDNN has seen experimental support for ARM64, POWER9, s390x, and even some level of NVIDIA GPU support.

The big addition with oneDNN 2.0 is now officially supporting Intel integrated and discrete graphics via oneAPI Level Zero. There is also Intel Data Parallel C++ (DPC++) compiler support as part of this release. With the Intel graphics support is also a unified shared memory (USM) implementation.

In the coming weeks I'll be working on some oneDNN 2.0 benchmarks atop Intel graphics as well as a fresh round of CPU testing. Phoronix Test Suite tests of oneDNN 1.x can be found via the OpenBenchmarking.org test profile page. I'll be updating the PTS test profile shortly for the new oneDNN 2.0 release.

More details on the oneDNN 2.0 release along with various Linux/Windows binaries via GitHub.

This comes with Intel on Tuesday announcing the 2021.1 oneAPI toolkits.