SYCL for CPU+GPU Heterogeneous Computing: A Study on Integrated and Discrete GPUs

This work presents a study on the influence of GPU capabilities on heterogeneous CPU+GPU computing schemes with SYCL. To conduct this analysis, two iterative problems were considered as case studies using dynamic load balancing between host and device. Performance evaluation was carried out on an integrated GPU (iGPU) as well as on two models of NVIDIA discrete GPUs (dGPUs). On the iGPU, shared memory with the CPU limits performance in memory-bound scenarios. In contrast, dGPUs consistently outperform the CPU due to their highly parallel architecture and high-bandwidth memory. Nonetheless, heterogeneous schemes may provide some benefits depending on the performance gap between devices and the computational characteristics of the problem.