First PCIe A100 performance test!?

We received the PCIe version of  Nvidia's latest GPU, A100, yesterday, so we decided to run some basic tests with our Falcon 4010, and this might be the first available test result for the PCIe A100 GPU you could find on the internet.

Equipment and System Setup

We have two servers, one is equipped with Intel Xeon Gold 5120T CPU, which only supports PCIe 3.0. The other one is equipped with AMD EPYC 7452, which supports PCIe 4.0. We installed the A100 GPUs in our Falcon 4010 expansion chassis, and connected the chassis to the two server hosts. The image below illustrates the system topology.

Methodology

To test how Falcon 4010 influences GPU peer to peer performance, we connect an Intel® Xeon® Gold 5120T PCIe Gen 3 server and an AMD EPYC™ 7452 PCIe Gen 4 server to the same Falcon 4010 expansion chassis. Each server is assigned four PCIe-A100 GPUs. We run CUDA Samples commands p2pBandwidthLatencyTest, htod, and dtoh on both host servers and compare the results.

Test Results Summary

Host to Device and Device to Host Bandwidth Performance

GPU Peer to Peer Bandwidth Performance

GPU Peer to Peer Latency

Overall, the bandwidth performance is pretty much as expected as PCIe 4.0 has doubled the bandwidth when compared to PCIe 3.0. However, you can see that the GPU peer to peer performance exceeds what it is supposed to be for the usual PCIe 3.0 environment. The result indicates that with Falcon 4010 (a PCIe 4.0 expansion chassis), you can still enjoy the PCIe 4.0 performance even when your CPU does not support PCIe 4.0.