high performance computing on graphics processing units: hgpu.org

For the first time in over 20 years, Intel is selling discrete graphics cards, including products for high-performance computing, scheduled for release in 2022. This thesis investigates programming models for the upcoming Intel GPUs and selects the Sycl standard, provided by oneAPI and hipSYCL, to port the heterogeneous components of SeisSol. The modules in question in SeisSol are analyzed and extended, and their efficiency is assessed using a Roofline Model analysis. Similar experiments are performed using Sycl with CUDA as a backend to test whether this combination can replace all native GPU approaches. This work demonstrates that already the low-power series of Intel’s discrete GPUs can reach over 85% of their peak performance using this portability, making them superior to integrated graphics chips. However, further measurements also show that Intel MPI needs to be improved before it can be used on supercomputers. With an Nvidia RTX 3090 and Intel Arria 10 FPGA, it is proven that Sycl’s cross-platform approach works but that its efficiency heavily depends on the implementation of the standard.