Characterizing Performance Improvement of GPUs
TimeTuesday, July 306:30pm - 8:30pm
LocationCrystal Foyer and Crystal B
DescriptionEvaluating supercomputer performance through benchmarking is an integral part of an HPC center operation. It can be used to guide the selection of a system during hardware acquisition, to validate that the selected hardware performs as expected, and to ensure that the system performs reasonably throughout its lifetime. In this work we quantify the performance improvement of our recent GPU architectures, GeForce RTX 2080 Ti, Titan V, and Tesla V100 over an older architecture, GeForce GTX 1080 Ti. We perform a workload analysis to select a suite of applications to benchmark. Using the benchmark results, a single performance number is then assigned to each GPU system under test. Taking hardware price into consideration, our results show consumer-grade GeForce RTX 2080 Ti as the most cost competitive GPU platform.