Pushing onwards from our recent addition of AMD accelerator support in OpenLB Release 1.9, the developer team is happy to share that we now have preliminary support for Intel GPUs. This means OpenLB now supports hardware acceleration across all of the big three platforms: NVIDIA, AMD, and Intel!
Using a new SYCL-based backend for our platform-transparent model implementations, we scaled up to a problem size of 4000 billion single-precision D3Q19 cells. Utilizing 1,000 nodes (~10%) of the Aurora supercomputer at the Argonne Leadership Computing Facility (TOP500 #3), this yielded a peak performance of 21,120 billion cell updates per second (GLUPs).
This work was done in the context of our ALCF Directorโs discretionary allocation project ๐๐ฑ๐ฆ๐ฏ๐๐ ๐ฐ๐ฏ ๐๐น๐ข๐ด๐ค๐ข๐ญ๐ฆ: ๐๐น๐ต๐ฆ๐ฏ๐ฅ๐ช๐ฏ๐จ ๐ฑ๐ญ๐ข๐ต๐ง๐ฐ๐ณ๐ฎ-๐ต๐ณ๐ข๐ฏ๐ด๐ฑ๐ข๐ณ๐ฆ๐ฏ๐ค๐บ ๐ต๐ฐ ๐๐ฏ๐ต๐ฆ๐ญ ๐๐ฆ-๐๐๐ (OLEX).
Interested in this or any other aspect of OpenLB and Lattice Boltzmann Methods? Join our upcoming Spring School this March in Liverpool, UK! (Early bird registration ends today)
