Brief performance portability analysis of a matrix multiplication kernel on multiple vendor GPUs

Abstract

The heterogeneous computing paradigm has led to the need for portable and efficient programming solutions that can leverage the capabilities of various hardware devices, such as NVIDIA, Intel, and AMD GPUs. This study evaluates the performance and portability of the SYCL and CUDA languages for a matrix multiplication (MM) application across different GPU architectures. The experimental work showed that, while the CUDA implementation outperforms the SYCL implementation on NVIDIA devices due to optimizations provided by the nvcc compiler, the latter implementation demonstrated remarkable code portability to other GPU architectures, such as AMD and Intel. Furthermore, the architectural efficiency percentages obtained on AMD and Intel GPUs showed consistency with the results observed on NVIDIA devices.