Impact of Memory Bandwidth on the Performance of Accelerators
Sambit Mishra, Dhruva K. Chakravorty, Lisa M. Pérez, Francis Dang, Honggao Liu, Freddie Witherden · 2024
This study investigates the impact of memory bandwidth of accelerators on the performance of computational simulations, revealing the importance of bandwidth over computational power in scalable high-order numerical simulations. A detailed analysis performed on an NVIDIA H100 GPU and an Intel MAX 1100 GPU on the NSF ACES platform, demonstrates how matrix multiplication characteristics such as matrix size and sparsity influence the demand for memory bandwidth. Utilizing the open-source fluid flow solver PyFR for the study for its flexibility, efficiency, and alignment with expected performance, this work emphasizes the necessity for accelerator designs to prioritize memory bandwidth to enhance simulation efficiency, particularly in the case of workloads whose performance is bound by available memory bandwidth.