SYCL for HPC: Adapting to Diverse CPU Architecture

Heterogeneous systems are prominent in High Performance Computing (HPC), offering a proven path to achieve higher FLOPS with fewer nodes and reduced power consumption. To fully exploit the potential of these systems, multiple parallel computing models emerged, targeting specific architectures. This emphasized the necessity of a unified programming model capable of executing across all architectures. SYCL, an open-source standard is gaining traction among the HPC community due to its single source programming approach, use of modern C++ and portability to any architecture. The paper presents our implementation of SYCL, designed to support X86, ARM and RISC based CPU architectures. To evaluate the effectiveness of the proposed implementation, several key metrics are investigated, including application scaling, performance, scalability, versatility, and software portability while covering the computational workloads based on Berkeley dwarfs. This work contributes to the ongoing efforts of facilitating broader adoption of unified programming models in HPC.