Our MLPerf training submission v2.0 includes multinode and single-node SUT.

The following general syntax is used for the system name: 

<Number of servers> x < Dell server name> x <number of accelerators> x <NVIDIA accelerator name> 

To identify a single-node system, note that there are no entries for the number of servers or the number of servers is equal to one. For example, the following are single-node systems:

• R750xax4A100-PCIE-80GB
• R750xax4A100-PCIE-80GB-NVBRIDGE
• XE8545x4A100-SXM-40GB
• XE8545x4A100-SXM-80GB
• DSS8440x8xA30-NVBRIDGE

The following figure shows the performance of different single node systems:

Figure 12.    Performance of single node systems on ResNet50 with NVIDIA A100 and A30 accelerators 

The time to converge with the NVIDIA 4xA100-PCIe-80GB accelerator is less than the NVIDIA 4xA100-SXM-40GB accelerator with the ResNet50 model, as shown in the following figures: 

Figure 13.    Performance of single-node systems on BERT with NVIDIA A100 and A30 accelerators 

Figure 14.    Performance of single-node systems on RNN-T with NVIDIA A100 and A30 accelerators 

Figure 15.    Performance of single-node systems on RetinaNet with NVIDIA A100 and A30 accelerators 

The following figure compares the performance of the NVIDIA A30 accelerator with the NVIDIA A100 accelerator:

Figure 16.     8xA30 compared to 4xA100 performance

The previous graphs show that single-node PowerEdge R750xa, Dell DSS 8440, and PowerEdge XE8545 systems performed well with NVIDIA A100 and A30 accelerators across different workload categories. The NVIDIA A100-SXM-80GB accelerator with the PowerEdge XE8545 server delivers the fastest time to value across different workload types.