With this latest release of oneAPI, we are continuing to improve stability and performance, with significant updates to both our plugins for NVIDIA and AMD GPUs. These plugins require the Intel oneAPI Base Toolkit, and when used together enable heterogeneous code that works across Intel, NVIDIA and AMD GPUs - for faster innovation, and greater productivity.

For NVIDIA GPUs, we brought an experimental implementation of the SYCL Graph extension. This major addition enables the ability to efficiently record and replay graph commands - reducing overheads and creating the potential for significant performance improvements. As this remains experimental, we welcome and encourage any feedback on this through our forums. If you are interested in reading more about SYCL Graph, this deep-dive blog was prepared by our Senior Software Engineer Ben Tracy.

Alongside this important addition, we've resolved a number of bug issues that improve overall stability of the plugin.

Full changelog can be found here.

For AMD GPUs, we have made improvements to overall support and continue on the path out of beta, to full release, and have brought support for half floating point types. Support for the SYCL 2020 subgroup operations has also been improved.

Experimental support for AMD Matrix Cores on CDNA2 architectures is here, via the joint matrix SYCL extension - which means that SYCL applications will be able to utilise Matrix Cores for more efficient matrix multiplication operations. Feedback on this experimental support is welcomed through our discourse forums!

Finally, general stability and performance improvements have been brought through bug-fixes.

Full changelog can be found here.