NVIDIA HGX B200 Specs & Datasheet (8-GPU Node)

The HGX B200 is NVIDIA’s 8-GPU Blackwell server platform — the current mainstream data-center node for large-model training and high-throughput inference. Eight HGX B200 GPUs sit on a single fifth-generation NVLink baseboard, so the node behaves as one tightly-coupled 8-GPU accelerator rather than eight discrete cards.

Unlike a rack-scale system such as the GB300 NVL72, the HGX B200 is a node you rack in a standard data center: a single 4U–10U chassis (depending on integrator and cooling) that drops into a conventional rack. It is the Blackwell successor to the Hopper HGX H100/H200 8-GPU nodes, and it is offered by multiple OEMs in both air- and liquid-cooled builds.

The HGX B200 8-GPU node specification, aggregated from the current integrator builds (Lenovo, Dell, Supermicro, Gigabyte). The per-GPU and per-node GPU, memory, and NVLink figures below are fixed characteristics of the platform; chassis, CPU, cooling, site power draw, and final configuration vary by integrator and are confirmed at quote.

Each B200 GPU in the HGX B200 carries 180 GB of HBM3e — the on-package GPU memory (VRAM) available to the model. Across the 8-GPU baseboard that is 1,440 GB of HBM3e per node, pooled and addressable over NVLink so a large model and its long-context KV cache can span all eight GPUs.

Memory bandwidth is roughly ~8 TB/s per GPU (derived from the ~64 TB/s aggregate per-node figure the integrators publish, divided across eight GPUs — treat it as approximate), for about ~64 TB/s aggregate across the node. That combination — 1,440 GB of high-bandwidth memory fed at ~64 TB/s — is what lets the HGX B200 keep more of a frontier-scale model resident and well-fed than the prior Hopper generation.

On compute, each B200 delivers ~4.5 PFLOPS of dense FP8 and ~9 PFLOPS of dense FP4 tensor throughput — about 36 and 72 PFLOPS respectively across the 8-GPU node — with the FP4 path aimed at high-throughput inference. All figures are dense; NVIDIA’s DGX B200 datasheet quotes the node totals "with sparsity" (72 PFLOPS FP8 / 144 PFLOPS FP4), i.e. 2× the dense figures. (The GB200 Grace-Blackwell superchip runs a higher clock bin at ~5 / ~10 PFLOPS dense per GPU.)

Inside the node, the eight B200 GPUs are wired over a fifth-generation NVLink baseboard at 1.8 TB/s per GPU — double the per-GPU NVLink bandwidth of the Hopper generation — giving all-to-all connectivity across the 8-GPU domain with minimal communication overhead.

To scale beyond a single node, the HGX B200 uses NVIDIA ConnectX-7 adapters at 400 Gb/s each with GPUDirect — eight per node, for roughly ~3.2 Tb/s of scale-out bandwidth per node on InfiniBand or Ethernet fabrics. (Supermicro also offers a higher-bandwidth ConnectX-8 build on its liquid-cooled 4U variant.) NVLink handles the dense traffic inside the node; ConnectX stitches many nodes into a larger training cluster.

The HGX B200 ships in two thermal designs, and the choice drives the form factor:

• Air-cooled (6U–10U). The mainstream build — the Dell PowerEdge XE9680 at 6U, the Supermicro 8U, and the Gigabyte 10U chassis — drops into standard data-center racks with no liquid loop, the fastest facilities fit.
• Liquid-cooled (4U–5U). A denser, direct-to-chip liquid build — the Dell XE9685L and Supermicro’s DLC variant at 4U, or the Lenovo ThinkSystem SR780a V3 at 5U with Neptune cooling — packs the same 8 GPUs into fewer rack units for high-density deployments, but needs a cooling loop or CDU.

The air-vs-liquid trade-off is the same one that governs the Hopper HGX H100 (see HGX H100 air vs liquid cooling): liquid buys density at the cost of facility water and plumbing, while air is simpler to deploy. Both carry the identical GPU, memory, and NVLink specs — only density and facility requirements differ.

The HGX B200 ships as a factory-integrated 8-GPU node from several OEMs — Lenovo (ThinkSystem SR780a V3), Dell (PowerEdge XE9680 / XE9685L), Supermicro, and Gigabyte — each with its own chassis, cooling, host CPU, and support model. All are new and factory-integrated: the GPU, memory, and NVLink specifications are common across builds, while chassis, cooling, host CPU, warranty, and lead time differ by integrator. Each build below links to its full specification.

Every HGX B200 build is quoted per configuration on request — the figure depends on integrator, cooling (air 8U vs liquid 4U), host CPU, networking, and quantity. Lead times run from ready-to-ship to a few weeks by integrator, confirmed at quote. Tell us the build you are standing up and we will return pricing, availability, and a facility-fit review. Browse the GPU catalog to compare integrations, or see how the B200 stacks up against the Hopper HGX H200.