NVIDIA HGX B300 carries eight Blackwell Ultra GPUs with 2.1 TB of HBM3e and 62 TB/s of memory bandwidth in total, linked at 1.8 TB/s per GPU over NVLink; each GPU is configurable up to 1,100 W. A DGX B300 built on it draws about 14.5 kW in 10 rack units and is air-cooled. Some NVIDIA documents and several server makers' sheets list 288 GB per GPU and 2.3 TB per server instead of 270 GB and 2.1 TB.
B300 (Blackwell Ultra) is NVIDIA's data-center GPU after B200. For rental and purchase it comes as 8-GPU HGX B300 servers and NVIDIA's own DGX B300; the 72-GPU rack version is GB300 NVL72. Every figure below links to its source.
| Spec | B300 | Source |
|---|---|---|
| Chip | 208 billion transistors, TSMC 4NP, two dies joined at 10 TB/s | NVIDIA technical blog |
| Memory | 270 GB HBM3e at 7.7 TB/s; 288 GB in the DGX B300 user guide and in Dell and Lenovo sheets (Supermicro gives both) | Blackwell Ultra datasheet, DGX B300 user guide, Dell spec sheet, Lenovo Press, Supermicro |
| NVLink | 1.8 TB/s per GPU (18 links of 100 GB/s) | NVIDIA technical blog |
| Power | Configurable up to 1,100 W | Blackwell Ultra datasheet, Dell spec sheet, Lenovo Press |
| FP4 Tensor | 14 dense, 18 sparse | Blackwell Ultra datasheet |
| FP8 / FP6 Tensor | 4.5 dense, 9 sparse | Blackwell technical brief |
| FP16 / BF16 Tensor | 2.2 dense, 4.5 sparse | Blackwell technical brief |
| TF32 Tensor | 1.1 dense, 2.2 sparse | Blackwell technical brief |
| FP32 / FP64 | 75 TFLOPS / 1.2 TFLOPS | Blackwell Ultra datasheet |
Tensor figures are NVIDIA's, in petaFLOPS: dense first, then with sparsity, as NVIDIA states them. NVIDIA rounds per-GPU and per-system figures separately, so a system figure is not always eight times the GPU figure.
| Spec | HGX B300 | Source |
|---|---|---|
| GPU memory | 2.1 TB HBM3e, 62 TB/s; 2.3 TB in the DGX B300 user guide | NVIDIA HGX, Blackwell Ultra datasheet, DGX B300 user guide |
| NVLink | 14.4 TB/s in total | NVIDIA HGX |
| FP4 Tensor | 108 dense, 144 sparse | NVIDIA HGX |
| FP8 / FP6 Tensor | 72 sparse (dense is half) | NVIDIA HGX |
| Networking | 1.6 TB/s per system; one ConnectX-8 SuperNIC at 800 Gb/s per GPU | NVIDIA HGX, NVIDIA, 18 March 2025 |
| CPU | None on the HGX board; each server adds its own CPUs | Blackwell Ultra datasheet |
| Spec | DGX B300 | Source |
|---|---|---|
| CPUs | Two Intel Xeon 6776P, 64 cores each | NVIDIA DGX B300, Intel |
| System memory | 2 TB, configurable to 4 TB | DGX B300 datasheet |
| Networking | Eight ConnectX-8 at up to 800 Gb/s InfiniBand or Ethernet; two BlueField-3 DPUs at up to 400 Gb/s | NVIDIA DGX B300, DGX B300 datasheet |
| Power | 14.5 kW and 15 kW system maximum (user guide); 14.5 kW on busbar, 15.1 kW with power supplies (datasheet); about 14 kW (product page); estimated peak 19 kW on busbar, 19.7 kW on AC (data-center guide) | DGX B300 user guide, DGX B300 datasheet, NVIDIA DGX B300, NVIDIA data-center guide, DGX B300 |
| Size | 10 rack units; 442 x 482.6 x 904.2 mm in the user guide, 17.5 x 19.0 x 35.3 in in the data-center guide | DGX B300 user guide, NVIDIA data-center guide, DGX B300 |
| Weight | 168 kg with power supplies, 123 kg in the busbar version | DGX B300 user guide |
| Cooling | Air; up to about 2,145 CFM of supply air per system | NVIDIA data-center guide, DGX B300 |
| Inlet air | 10 to 30 °C in the user guide, 10 to 35 °C in the datasheet | DGX B300 user guide, DGX B300 datasheet |
NVIDIA's data-center guide for DGX B300 gives two rack layouts: four systems per rack on four 33 kW DC power shelves, about 58 kW on average and 76 kW at peak, cooled by air with an active rear-door heat exchanger (rack weight 1,122 kg); or two systems per rack on AC power, about 30 kW on average and 39.4 kW at peak. NVIDIA does not recommend passive rear-door heat exchangers for DGX B300 (NVIDIA data-center guide, DGX B300).
| Server | Cooling and size | Power supplies | Weight | Source |
|---|---|---|---|---|
| Supermicro SYS-822GS-NB3RT | Air, 8U | Six 6,600 W (3+3) | 119 kg | Supermicro |
| Supermicro SYS-422GS-NB3RT-LCC | Direct-to-chip liquid, 4U; up to 64 GPUs per rack | Four 6,600 W (2+2) | 84.6 kg | Supermicro, Supermicro, 9 December 2025 |
| Dell PowerEdge XE9780 | Air, 10U | Twelve 3,200 W | 163.2 kg | Dell spec sheet |
| Dell PowerEdge XE9780L | Liquid, 3 OU in an IR7000 rack | Six 5,500 W in the rack power shelf | 107.2 kg | Dell spec sheet |
| Lenovo ThinkSystem SR680a V4 | Air, 8U; up to five per rack | 3,200 W (N+1) or 3,800 W (N+N) | Up to 124.7 kg | Lenovo Press |
| Gigabyte G894-SD3-AAX7 | Air, 8U | Twelve 3,000 W | Gigabyte datasheet |
In our offers reserved HGX B300 can start from Q4 2026, at about $5.40 per GPU-hour at the median for two years or longer, our fee included. See B300 rental, what a B300 server costs to buy and B300 vs GB300 NVL72.
270 GB of HBM3e per GPU in NVIDIA's Blackwell Ultra datasheet, 2.1 TB per 8-GPU HGX B300. The DGX B300 user guide and several server makers' sheets list 288 GB per GPU and 2.3 TB per server.
Each B300 GPU is configurable up to 1,100 W (NVIDIA; Dell and Lenovo state the same).
14.5 kW in NVIDIA's user guide (15 kW system maximum), 14.5 to 15.1 kW in the datasheet and about 14 kW on the product page, with an estimated peak of 19 kW on busbar power.
Both exist: DGX B300 and 8U or 10U servers from Supermicro, Dell, Lenovo and Gigabyte are air-cooled; 4U Supermicro and Dell XE9780L servers use direct-to-chip liquid cooling. GB300 NVL72 racks are liquid-cooled.
NVIDIA announced it on 18 March 2025, and Supermicro announced volume shipments of HGX B300 systems on 11 September 2025.
