The storage performance target for training or inference can vary depending on the type of model and dataset. The guidelines in Table 1 provide standard throughput for the various Scalable Unit (SU) system sizes. The final HPS requirements for throughput and capacity will be specified for each NCP opportunity.
| Compute Components |
Performance Guidance (GB/s) |
| GPUs |
Nodes |
SUs |
Read |
Write |
| 1,024 |
128 |
4 |
160 |
80 |
| 2,048 |
256 |
8 |
320 |
160 |
| 8,192 |
1,024 |
32 |
1,280 |
640 |
| 16,384 |
2,048 |
64 |
2,560 |
1,280 |
Table 1.
NVIDIA HGX standard performance guidelines for NCPs