Feature summary
| GPU | 8 x NVIDIA B300 Blackwell Ultra GPUs |
|---|---|
| GPU memory | 8 x 288 GB = 2.3 TB total |
| Performance | 72 PFLOPS FP8 training · 144 PFLOPS FP4 inference |
| NVSwitch | 2 x fifth-generation NVLink interconnects |
| CPUs | 2 x Intel Xeon Platinum 6776P |
| System memory | 2 TB default, up to 4 TB with 128 GB DIMMs from NVIDIA, configurable on site |
| Cluster network | 8 x OSFP ports to 8 x ConnectX-8 cards · 8 x 800 Gb/s InfiniBand or Ethernet |
| Storage & management network | 2 x dual-port BlueField-3 DPUs · 2 x 400 Gb/s InfiniBand or Ethernet |
| Cache storage | 8 x E1.S 3.84 TB NVMe self-encrypting drives |
| Boot storage | 2 x 1.92 TB M.2 NVMe, software encryptable |
| Host management | On-board 1 GbE RJ-45 |
| Remote management | BMC on a replaceable DC-SCM · 1 GbE RJ-45 · remote KVM & storage · Redfish & IPMI |
| Operating system | DGX OS 7 on Ubuntu 24.04 LTS; Ubuntu, Red Hat Enterprise Linux 8 & 9, Rocky Linux also supported |
Physical & operational
| Rack units | 10 RU |
|---|---|
| System size (H x W x L) | 17.4 x 19 x 35.74 in (442 x 482.6 x 904.2 mm) |
| System weight | PSU version 370.38 lb (168 kg) · busbar version 271.17 lb (123 kg) |
| Packaged | PSU 457.68 lb (207.6 kg) · busbar 358.47 lb (162.6 kg) · 41.93 x 28 x 46.38 in |
| Power consumption | 14.5 kW |
| Power supplies | 12, N+N, front-loaded with rear connections. The user guide rates each supply 3.2 kW in its specifications table and 3.3 kW in its power-specifications table; the system maximum is 15 kW either way |
| Power inlets | 12 (AC version only) |
| Voltage | PSU 200 to 240 VAC · busbar 54 VDC |
| Max current per outlet | PSU 16 A · busbar 300 A |
| Heat output | 49,476 BTU/hr max |
| Airflow at 70% PWM | PSU 1500 CFM · busbar 1350 CFM |
| Inlet air temperature | 10 to 30 C (50 to 86 F) |
Power & redundancy
The system ships in two power versions. The PSU version takes twelve power supplies for N+N redundancy: six carry the load and six are spare, the minimum for operation is six, and the system is capped at 15 kW, so the supplies' combined rating is not the number to multiply. NVIDIA's user guide gives two ratings for one supply, 3.2 kW in its specifications table and 3.3 kW at 200 to 240 V, 16 A, 50 to 60 Hz in its power-specifications table; both are quoted here rather than reconciled, and the guide itself is the document to cite. The busbar version has a rear clip for a 54 VDC busbar at 300 A maximum; the power shelves that energise the busbar are not included with the node.
- In N+N, the minimum number of PSUs is six.
- The node boots only when at least three PSUs are operational; with fewer, only the BMC is available.
- A failed PSU should be replaced immediately; the node is never operated with PSUs depopulated.
- Twelve locking cords ship with the node: 250 VAC, 20 A, C19/C20, 1200 mm, UL62 / IEC60227 cord, IEC60320-1 connector. NVIDIA specifies only these cords are used.
What changed from earlier DGX systems
- AC/PDU & DC/busbar versions.
- All I/O moved to the front; ConnectX-8 connects to the GPUs directly over PCIe Gen6, with the NICs & cages housed in the GPU tray.
- Front-loaded power supplies with rear connections; all fans accessible from the rear.
- E1.S cache drives replace U.2 at the same capacity & configuration.
- Motherboard & GPU trays on sliding rails; a replaceable DC-SCM carries the BMC.
- A black front with a gold bezel, which is where our brass stripe comes from.
Front panel
| Power button | Green flashing 1 Hz: standby, BMC booted · green flashing 4 Hz: POST in progress · green solid: powered on |
|---|---|
| ID button | Blue LED for servicing, also flashes an LED at the rear; configurable from the BMC |
| Fault LED | Amber: system or component fault |
| BMC port LEDs | Left: solid green 100M, solid amber 1G · right: green flashing on activity |
Software stack as shipped
DGX OS on an Ubuntu server distribution with the optimised kernel; NVIDIA System Management (NVSM) for health & alerts; Data Center GPU Management (DCGM) for node-wide GPU administration; the NVIDIA GPU driver with CUDA; Docker Engine; the NVIDIA Container Toolkit; DOCA-OFED; the networking software tools (MST); cachefilesd.
Discrepancy note: NVIDIA's product page has been reported to list 2.1 TB of GPU memory where the user guide & datasheet list 2.3 TB (8 x 288 GB). This site uses the user-guide figure throughout.
