A Comparative Analysis of Kunpeng 920 on HPC Workloads
Yinchuan Guo, Xianwei Zhang · 2024
As ARM architecture gains traction in server usage, the Kunpeng 920 processor, developed by HiSilicon specifically for cloud computing, has emerged as a notable contender. With its impressive configuration of 64 cores and a bandwidth of 200 GB/s, Kunpeng 920 delivers formidable computational capability. However, its suitability for HPC applications remains relatively unexplored. This study conducts a comprehensive evaluation of the Kunpeng 920, comparing its performance to AMD EPYC 7302 across various HPC workloads. We delve into its computational capabilities, memory subsystem efficiency, and performance under real-world workloads. Employing methodologies like the roofline model and Top-down Microarchitecture Analysis (TMA), we identify potential bottlenecks for HPC applications running on the Kunpeng 920. We reveals that Kunpeng 920 excels in memory-intensive tasks due to its superior memory bandwidth, while gaining poor performance in compute-intensive tasks. Furthermore, this research uncovers previously unrecognized issues with the TMA model’s definition for Kunpeng processors.