Memory tricks
- DCGM diagnostic levels: 1 = short, 2 = medium, 3 = long (small → big)
- MIG profile format:
Xg.Ymb— X GPU slices, Y GB memory - Fleet Command = Edge (not data center); BCM = Data center clusters
- Fabric Manager = NVLink/NVSwitch; UFM = InfiniBand — never mix these
- DOCA = BlueField DPU software (not GPU software)
- GDS = storage → GPU (skip the CPU)
- NGC login: username is always
$oauthtoken, password is your API key
