NVIDIA DGX GB300
A liquid-cooled, rack-scale DGX that links 72 Blackwell Ultra GPUs and 36 Grace CPUs in one NVLink domain. Built for enterprises and AI factories that want NVIDIA's own turnkey reasoning and training system.
Facts
Source per value- Maker The OEM, ODM or system builder that sells the server model. Corporate names follow the vendor’s own pages.
- NVIDIA 1
- GPU vendor Maker of the accelerators named in the GPU model field: NVIDIA, AMD or Intel. Blank when the vendor page names no accelerator.
- NVIDIA 1
- GPU or accelerator Accelerators the vendor page names for this model. A model that supports several options lists the option that was read, marked as one supported option.
- NVIDIA Blackwell Ultra (GB300 Grace Blackwell Ultra Superchips) 1
- GPUs per server GPUs Installed count or supported maximum per server, as the qualifier in the source says. Rack-scale systems show the GPUs in the whole rack (72 for an NVL72). Blank when the page gives no count.
- 72 1
- Chassis height U Rack height of the server as the vendor states it. Open Rack Units (OU) are not the same as EIA U and sit in the form factor field instead.
- Not published
- Form factor Configuration-specific chassis or rack description where the vendor splits it by cooling or rack standard.
- Rack-scale, NVIDIA MGX rack 1
- Cooling type Air, liquid, or air or liquid by configuration, read from the vendor’s cooling statement. Blank when the page does not state cooling.
- Liquid 1
- Cooling (vendor wording) The vendor’s cooling statement with configuration qualifiers kept.
- Rack-scale, liquid-cooled system 1
- CPU support Processor families or installed processors the named model supports.
- 36 x NVIDIA Grace CPU (2,592 Arm Neoverse V2 cores) 1
- CPU sockets sockets Processors per server.
- Not published
- CPUs per rack CPUs Total CPUs in a rack-scale configuration.
- 36 1
- Memory per GPU GB Memory on each accelerator in the cited configuration.
- Not published
- Total GPU memory GB Vendor-stated total accelerator memory per system.
- Not published
- Max system memory TB Vendor-stated maximum host RAM, not GPU memory.
- Not published
- Memory slots DIMMs Host memory DIMM slots.
- Not published
- GPU TDP W Thermal design power per GPU, not measured server draw.
- Not published
- System power kW Maximum power the vendor states for one system, as printed on its own page.
- Not published
- Rack power kW Power the vendor states for a full rack-scale system.
- Not published
- Power supplies PSU ratings, quantity and redundancy where published. Rating, not draw.
- Not published
- Rack power delivery Power-shelf rating in rack-scale and rack-sled systems. Delivery capacity, not rack demand.
- Not published
- GPU interconnect GPU fabric and published link bandwidth.
- NVIDIA NVLink Switch System with 9 x L1 NVLink Switches 1
- Networking Published ports, adapters and speeds.
- 72 x OSFP single-port ConnectX-8 VPI (800 Gb/s InfiniBand); 18 x dual-port BlueField-3 VPI (200 Gb/s InfiniBand and Ethernet) 1
- Source date Publication date printed on the cited page, where the vendor prints one.
- Not published
- Warranty Base warranty or service statement the vendor publishes.
- Not published
- Operating temperature °C Supported ambient range with the vendor’s caveats.
- Not published
- Maker headquarters Headquarters or principal corporate office of the maker, not a factory location.
- Santa Clara, United States 2
Sources
- www.nvidia.com accessed 2026-11-11
- www.nvidia.com accessed 2026-11-11
Related servers
In these lists
Get a quote from NVIDIA DGX GB300 and similar servers
Tell us what you need once. We send it to matching servers and they reply to you directly.