We used two workloads to conduct the performance studies:
- ResNet inference—We used ResNet50 version 1.5, which is available as a container in NVIDIA NGC. This workload runs inference on a dataset of images.
- ResNet training—We used ResNet50. This workload runs ResNet50 model training on a dataset of images.
We used ‘images per second’ as the metric to measure performance. For the training runs, the number of CPU threads were limited to 16 for both VM and bare-metal configurations. We performed three training runs. The results presented are an average of these three runs.
The following table shows the bare-metal server configuration:
Table 7. PowerEdge R740xd server configuration
CPU | Intel Xeon Gold 6148 CPU 20 cores @ 2.40 GHz |
Memory | 256 GB |
GPU | NVIDIA A100-PCIE-40 GB |
NIC | ConnectX-5 DX 25 GbE |
Storage | PowerStore Isilon H400 |
Software (vSphere server only; not applicable for bare metal) | VMware vSphere 7.0 U2 |
The following table shows the VM server configuration:
Table 8. VM configuration
CPU | 16 |
Memory | 64 |
Storage | 500 GB thin-provisioned virtual disk on Isilon storage |
We enabled MIG in the ESXi server for all the training runs.
