This section describes performance improvements from MLPerf Training v1.1 to MLPerf Training v2.0.
The following figures show the performance improvements from the same system (having identical servers and accelerators):
Figure 21. Performance improvements between PCIe and SXM form factor GPUs for ResNet on PowerEdge R750xa and XE8545 servers respectively
The performance gains seen with the PowerEdge R750xa server with four A100-PCIE-80GB cards are 4.54 percent and with two PowerEdge XE8545 servers with four A100-SXM-80GB cards are 4.93 percent for the ResNet50 model.
Figure 22. Performance improvements between PCIe and SXM form factor GPUs for BERT on PowerEdge R750xa and XE8545 servers respectively
The performance gains seen with the PowerEdge R750xa server with four A100-PCIE-80GB cards are 3.55 percent and with two PowerEdge XE8545 servers with four A100-SXM-80GB cards are 9.30 percent for the BERT model.
The performance gains come from software-only improvements with the same hardware and most of the gains are from the SXM form factor card.