NVIDIA HGX B200 carries eight Blackwell GPUs with 1.4 TB of HBM3e in total, linked at 1.8 TB/s per GPU over NVLink; each GPU is configurable up to 1,000 W. A DGX B200 built on it draws up to 14.3 kW in 10 rack units and is air-cooled. NVIDIA lists up to 192 GB per GPU; servers from Dell, Lenovo and Supermicro ship 180 GB per GPU.
B200 is the first Blackwell data-center GPU. It comes as 8-GPU HGX B200 servers and NVIDIA's DGX B200; the 72-GPU rack version is GB200 NVL72. Every figure below links to its source.
| Spec | B200 | Source |
|---|---|---|
| Memory | Up to 192 GB HBM3e at 7.7 TB/s (NVIDIA technical brief); 180 GB per GPU in Dell, Lenovo and Supermicro servers, and 1,440 GB per 8-GPU DGX B200 in NVIDIA's user guide | Blackwell technical brief, Lenovo Press, Dell spec sheet, DGX B200 user guide |
| NVLink | 1.8 TB/s per GPU | NVIDIA HGX |
| Power | Configurable up to 1,000 W | Blackwell technical brief, Lenovo Press, Dell spec sheet |
| FP4 Tensor | 9 dense, 18 sparse | Blackwell technical brief |
| 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 / 37 TFLOPS | Blackwell technical brief |
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 B200 | Source |
|---|---|---|
| GPU memory | 1.4 TB HBM3e; up to 62 TB/s (64 TB/s in the DGX B200 datasheet) | NVIDIA HGX, Blackwell technical brief, NVIDIA DGX B200 |
| NVLink | 14.4 TB/s in total | NVIDIA HGX |
| FP4 Tensor | 72 dense, 144 sparse | NVIDIA HGX |
| FP8 / FP6 Tensor | 36 dense, 72 sparse | NVIDIA DGX B200 |
| FP64 | 296 TFLOPS | NVIDIA HGX |
| Networking | 0.8 TB/s per system | NVIDIA HGX |
| Spec | DGX B200 | Source |
|---|---|---|
| CPUs | Two Intel Xeon Platinum 8570, 112 cores in total | NVIDIA DGX B200 |
| System memory | 2 TB, configurable to 4 TB | NVIDIA DGX B200 |
| Networking | Eight ConnectX-7 at up to 400 Gb/s InfiniBand or Ethernet; two BlueField-3 DPUs | NVIDIA DGX B200 |
| Power | 14.3 kW maximum; six 3.3 kW power supplies, 5+1 redundant | DGX B200 user guide |
| Size | 10 rack units, 444 x 482.2 x 897.1 mm | NVIDIA DGX B200 |
| Weight | 142.4 kg maximum | DGX B200 user guide |
| Cooling | Air; 1,550 CFM airflow in the user guide, up to about 2,145 CFM of supply air in the data-center guide | DGX B200 user guide, NVIDIA data-center guide, DGX B200 |
NVIDIA's data-center guide for DGX B200 plans two systems per rack as the typical layout, at 28.6 kW, or four per 52U rack at 57.2 kW, and notes that rear-door heat exchangers and in-row coolers are not typically suitable for DGX B200 (NVIDIA data-center guide, DGX B200).
| Server | Cooling and size | Power supplies | Source |
|---|---|---|---|
| Supermicro SYS-A21GE-NBRT | Air, 10U | Six 5,250 W (3+3) | Supermicro |
| Supermicro SYS-422GA-NBRT-LCC | Direct-to-chip liquid, 4U; 92 kg | Four 6,600 W (2+2) | Supermicro |
| Dell PowerEdge XE9680L | Liquid, 4U | 3,000 W units | Dell spec sheet |
| Lenovo ThinkSystem SR680a V3 / SR780a V3 | Air / water-cooled (Lenovo Neptune) | Lenovo Press |
In our offers B200 can start from Q4 2026. See B200 rental, what a B200 server costs and B200 vs B300.
Up to 192 GB of HBM3e per GPU in NVIDIA's technical brief; HGX B200 servers from Dell, Lenovo and Supermicro ship 180 GB per GPU, and NVIDIA's DGX B200 user guide gives 1,440 GB for its eight GPUs.
Each B200 GPU is configurable up to 1,000 W in HGX B200 and DGX B200.
14.3 kW at most, from six 3.3 kW power supplies (NVIDIA user guide).
DGX B200 and 10U OEM servers are air-cooled; 4U servers from Supermicro and Dell use direct-to-chip liquid cooling, and Lenovo offers a water-cooled version.
