Click a node to drill in, or click any component to learn about it.
This cluster is modeled on the NVIDIA DGX SuperPOD — the reference architecture for large-scale AI training. Each of the 4 DGX A100 nodes houses 8 A100 GPUs connected in a full NVLink mesh via NVSwitch. Nodes connect to each other via InfiniBand HDR (200 Gb/s per link). A real SuperPOD has 20 nodes; we show 4 here for clarity. Training GPT-3 (175B parameters) required clusters of this scale running for weeks.