You're reading the specs for a DGX A100 system and you see it includes multiple high-speed network interface cards. Specifically, how many Mellanox ConnectX-6 Single Port VPI cards are integrated into the system?
Select an answer to reveal the explanation.
Short Explanation
Here's an easy way to remember it: a DGX A100 has eight A100 GPUs, and it gives each one its own single-port ConnectX-6 card for the cluster fabric. That's 8 cards, not 4 — the separate dual-port card is for storage.
Full Explanation
The DGX A100 pairs each of its eight A100 GPUs with a dedicated network adapter, so the system ships with eight single-port Mellanox ConnectX-6 VPI cards for the compute cluster fabric, each running up to 200 Gb/s HDR InfiniBand. That one-to-one GPU-to-NIC layout is what lets GPUDirect RDMA traffic leave each GPU through a nearby adapter without crossing the CPU, giving roughly 1.6 Tb/s of aggregate cluster bandwidth for multi-node training. VPI means Virtual Protocol Interconnect: each card can run InfiniBand or Ethernet depending on configuration. Separate from those, the system carries dual-port ConnectX-6 VPI adapters for the storage and in-band management network, one in the base configuration and a second optional one on larger models. Four is wrong because it would leave half the GPUs without a local cluster adapter. Two confuses the count with the dual-port storage adapters. Sixteen does not match any DGX A100 configuration and overcounts the cluster cards. On the exam, tie the number to the GPU count: eight GPUs, eight single-port compute NICs. To verify on a live system, run ibstat or ibdev2netdev and count the single-port HDR devices, then compare the result against the network port mapping in the DGX A100 user guide.