Targeting NVIDIA Devices with Intel® oneAPI and SYCL

NVIDIA® GPUs and CUDA programming language has remained the standard in HPC and AI workloads forming a closed ecosystem. The arrival of diverse classes of accelerators have opened new opportunities for innovation across multiple industries. We introduce oneAPI, an open, cross-industry, standards-based, unified, multiarchitecture, multi-vendor programming model. oneAPI uses SYCL, an openly spec’d alternative to CUDA that supports many classes of accelerators like CPUs, GPUs, and FPGAs without vendor lock-in. In this talk, we demonstrate how switching to SYCL and oneAPI using the Intel® oneAPI DPC++/C++ Compiler with the NVIDIA® plugin can deliver performance parity between SYCL and CUDA on NVIDIA® GPUs. Through a careful selection of representative algorithms widely used in bioinformatics, image processing, finances, machine learning and physics, it is shown that SYCL can provide comparable performance with respect to native CUDA.

Download Presentation