MP-STREAM: A Memory Performance Benchmark for Design Space Exploration on Heterogeneous HPC Devices

Sustained memory throughput is a key determinant of performance in HPC devices. Having an accurate estimate of this parameter is essential for manual or automated design space exploration for any HPC device. While there are benchmarks for measuring the sustained memory bandwidth for CPUs and GPUs, such a benchmark for FPGAs has been missing. We present MP-STREAM, an OpenCL-based synthetic micro-benchmark for measuring sustained memory bandwidth, optimized for FPGAs, but which can be used on multiple platforms. Our main contribution is the introduction of various generic as well as device-specific parameters that can be tuned to measure their effect on memory bandwidth. We present results of running our benchmark on a CPU, a GPU and two FPGA targets, and discuss our observations. The experiments underline the utility of our benchmark for optimizing HPC applications for FPGAs, and provide valuable optimization hints for FPGA programmers