Cisco UCS C845A M8 (MGX, PCIe)
8 × H200 NVL / L40S / RTX PRO 6000 (PCIe) per node, with 4 compute rails at 400G. That is 1.6 Tb/s of scale-out network per node.
- GPUs per node8 × H200 NVL / L40S / RTX PRO 6000 (PCIe)
- Compute rails4 × 400G
- Scale-out bandwidth1.6 Tb/s
- Storage links1 × 200G
- In-band links1 × 100G
- BMC ports1
- Rack units4U
- Typical power6.5 kW
- Maximum power7.5 kW
What this means for the fabric
4 compute NICs means 4 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 6.5 kW typical draw, this node fills a rack's power budget long before it fills the space. A 45 kW rack takes 6 of them on power alone, against 10 on rack units. Power is what actually sets the rack count.
Notes
Built on the NVIDIA MGX reference design. Verified: 4RU, two 5th Gen AMD EPYC, up to 8 TB DDR5-6400, two to eight PCIe GPUs (H200 NVL, L40S, RTX PRO 6000, or AMD MI210), with 8 x PCIe x16 GPU slots and 5 x PCIe x16 FHHL slots. Aimed at fine-tuning and inference rather than large-scale training. The rail count here is a slot-budget decision, not one-per-GPU — verify against your actual configuration.
Source: Cisco UCS C845A M8 data sheet and spec sheet, verified 2026-08. Rack units, CPU, and GPU/PCIe slot counts confirmed; NIC count and power are estimates. Needs verification 2026-08-06