large_scale_gpus

We investigate the influence of computational hardware topology on the throughput performance in terms of double precision floating point operations per second (FLOPS). Two commercially available GPU-accelerated compute nodes are compared using the multiplication of dense matrices as a compute-bound mathematical problem. We find that, the higher host-device memory bandwidths provided by NVLink-enabled CPUs (compared to PCIe connections) significantly improve the overall performance. We also compare our own implementation with an NVIDIA benchmark and report substantial speedups, especially for very large matrices.