Workshops Archive

Performance comparison and analysis of macroscopic cross section lookup kernel on Cerebras CS-3 and NVIDIA GPUs


Workshop: PMBS26: The 17th International Workshop on Performance Modeling, Benchmarking, and Simulation of High Performance Computer Systems

Authors: Reo Nagashima and Akeru Nakamura (Meiji University), John R. Tramm (Argonne National Laboratory), and Takaaki Miyajima (Meiji University)

Abstract: The Cerebras CS-3 integrates approximately 900,000 processing elements in a two-dimensional mesh and provides high memory bandwidth and an exceptionally high B/F ratio, making it attractive for memory-bound scientific workloads. We evaluate a computational kernel that abstracts the macroscopic cross section lookup in XSBench, a Monte Carlo particle transport mini-app, characterized by low arithmetic intensity and irregular random memory accesses. We measure maximum performance, strong scaling, and weak scaling on CS-3, analyze the factors contributing to its improvement over CS-2, and compare it with optimized implementations on NVIDIA GH200 and A100 GPUs. Without additional optimization, the CS-3 implementation achieves 9.44E+09 lookups/s and an 11% higher geometric-mean performance than CS-2. Under identical experimental conditions, it delivers 6.34× and 11.90× higher throughput than GH200 and A100, respectively. These results demonstrate the potential of CS-3 for irregular, memory-intensive lookup workloads. Our analysis also highlights how wafer-scale architecture benefits this challenging class of kernels.


Back to PMBS26: The 17th International Workshop on Performance Modeling, Benchmarking, and Simulation of High Performance Computer Systems Archive Listing

Back to Full Workshop Archive Listing