Evaluating the performance of CUDA Graphs in common GPGPU programming patterns

Abstract: CUDA Graphs is a relatively new and unexplored feature of CUDA that presents an alternative for submitting work to the GPU. This report aims to evaluate the performance benefits of CUDA Graphs. The means to this end was a benchmark suite containing CUDA applications that do not utilise CUDA Graphs. Its BFS application was modified to do so and compared against its original. The results showed an improvement of up to 14 % for medium-sized workloads, but no difference for large workloads and worse performance for small workloads. In conclusion, the performance gains are noticeable albeit only for certain types of workload.