The Blackwell frontier accelerator. The B200 is a dual-die GPU with 192GB of HBM3e at 8 TB/s and fifth-generation NVLink at 1.8 TB/s, delivering roughly 2.5× the FP8 training throughput of Hopper and native FP4 for inference — the platform for the largest training clusters, built around direct liquid cooling.
Key specifications for the NVIDIA B200 Tensor Core GPU. Hard specs mirror the live catalog; confirm exact revision and configuration at quote.
| Specification | B200 |
|---|---|
| Architecture | Blackwell (dual-die, TSMC 4NP) |
| GPUs per unit | 1 (8 per HGX B200 server) |
| Memory | 192 GB HBM3e |
| Memory bandwidth | 8 TB/s |
| FP8 tensor | ~9,000 TFLOPS (with sparsity) |
| FP16 / BF16 tensor | ~4,500 TFLOPS (with sparsity) |
| NVLink | 1.8 TB/s (5th-gen NVLink) |
| Board power (TDP) | 1000 W (up to ~1,200W DLC) |
| Form factor | SXM6 |
| Cooling | Direct-to-chip liquid (DLC) |
| Typical lead time | 10–14 wk |
| Indicative price | From $38,500 / unit |
The B200 is built for scale-out training and the most demanding inference. Two reticle-limited dies act as one GPU with 192GB of HBM3e, a second-generation Transformer Engine, and native FP4 that pushes inference throughput far beyond Hopper. That density comes with heat: the B200 is a direct-liquid-cooled part, so it belongs in a DLC-ready hall with matched CDUs and power.
Roughly 2.5× the FP8 throughput of an H100 and 1.8 TB/s NVLink make it the engine for training the largest models on huge GPU counts.
Native FP4 and 192GB HBM3e give a step-change in tokens-per-second per rack for serving very large models cost-effectively.
Designed for direct-to-chip liquid cooling — quote the servers, CDUs and power train together so thermals and density line up from day one.
Density decides the facility. Size the electrical capacity, heat rejection and cooling method before you rack a single B200 — our free tools turn the specs above into facility numbers.
Add it to your quote list — free, no commitment — and we confirm a firm, itemized price, availability and lead time after review, usually within a business day.
The B200 is NVIDIA’s Blackwell-generation Tensor Core GPU — a dual-die accelerator with 192GB of HBM3e at 8 TB/s and fifth-generation NVLink at 1.8 TB/s. It roughly triples usable compute over an H100 and adds native FP4 for inference.
The B200 draws about 1,000W air-cooled and up to ~1,200W in its highest-performance direct-liquid-cooled configuration. At these densities direct-to-chip liquid cooling is the norm — plan a DLC-ready rack, matched CDU capacity and the power to feed it.
The B200 is a generational jump: about 2.5× the FP8 training throughput of Hopper, 192GB HBM3e at 8 TB/s (vs 141GB/4.8 TB/s on the H200 and 80GB/3.4 TB/s on the H100), and 1.8 TB/s NVLink. The trade-off is higher power and mandatory liquid cooling.
The B200 is a discrete GPU you rack in an HGX server. The GB200 is a superchip that pairs two B200 GPUs with a Grace CPU over NVLink-C2C; 72 of those Blackwell GPUs plus 36 Grace CPUs make up a GB200 NVL72 rack — a single 72-GPU NVLink domain.
Our solutions engineers size compute, power and cooling together and confirm real lead times — no payment, no commitment, quotes back in about one business day.
Our engineers scope power, cooling and compute together. No payment, and quotes come back in about a business day.