CERN* took advantage of the Intel® oneAPI model that supports multiarchitecture programming via C++ with SYCL* to make use of hardware accelerators like GPUs in the ATLAS software framework.
Intel® oneAPI Base Toolkit components helped accelerate and optimize the code blocks. The SYCL API made it possible to achieve asynchronous running of interdependent tasks. Yet, achieving the same for CUDA* code required extra effort.
- Intel® Data Center GPU Max Series 1550 performed in competition with an NVIDIA* A100 GPU.
- Performance of SYCL code on an NVIDIA back end was very close to native CUDA* code that runs the same algorithm.
- Intel Data Center GPUs performed significantly better than what would be possible using even multiple traditional CPUs.