NVIDIA DGX B300 (Spectrum-X Ethernet)
8 × B300 (Blackwell Ultra) per node, with 16 compute rails at 400G. That is 6.4 Tb/s of scale-out network per node.
- GPUs per node8 × B300 (Blackwell Ultra)
- Compute rails16 × 400G
- Scale-out bandwidth6.4 Tb/s
- Storage links2 × 400G
- In-band links2 × 100G
- BMC ports1
- Rack units10U
- Typical power13 kW
- Maximum power14 kW
What this means for the fabric
16 compute NICs means 16 rails. In a rail-optimized design each of those rails is sized on its own against the node count, then multiplied back up, so the leaf tier is driven by the rail count and not by the total endpoint count.
Storage and in-band management are separate fabrics with their own switches and their own optics, and the 1 BMC port per node feed the out-of-band network, which also has to carry every switch and every PDU in the cluster.
Power and racks
At 13 kW typical draw, this node fills a rack's power budget long before it fills the space. A 45 kW rack takes 3 of them on power alone, against 4 on rack units. Power is what actually sets the rack count.
Notes
The Ethernet face of the DGX B300, and materially different from the InfiniBand one. Verified: 64 nodes per scalable unit, 2x400GbE per GPU reaching two independent planes, 512 nodes in a full SuperPOD. Set planes to 2 in the compute fabric to match the reference design — leaving it at 1 sizes a single-plane fabric with 16 rails, which is a different topology.
Source: NVIDIA DGX SuperPOD B300 Spectrum-X reference architecture, verified 2026-08. Confirmed 2026-08-06