NVIDIA HGX B300 Specs & Datasheet (8-GPU Node)

The HGX B300 is NVIDIA’s 8-GPU Blackwell Ultra server platform — the memory-and-inference step-up over the HGX B200, and the current top-of-line 8-GPU node for the largest-memory training and high-throughput inference. Eight B300 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.

The HGX B300 is a node, not a rack. Do not confuse it with the GB300 NVL72 — that is a 72-GPU rack-scale system with a single 72-GPU NVLink domain. The HGX B300 is the discrete 8-GPU node you rack in a standard data center: a single chassis (air- or liquid-cooled, depending on integrator) that drops into a conventional rack. It is the Blackwell Ultra successor to the HGX B200, and it is offered by multiple OEMs in both air- and liquid-cooled builds.

The HGX B300 8-GPU node specification, aggregated from the current integrator builds (Supermicro, Dell, HPE, Lenovo). 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 B300 GPU in the HGX B300 carries 288 GB of HBM3e — the on-package GPU memory (VRAM) available to the model, a 60% step over the 180 GB on the HGX B200. Across the 8-GPU baseboard that is roughly 2,304 GB (~2.3 TB) of HBM3e per node, pooled and addressable over NVLink so a large model and its long-context KV cache can span all eight GPUs with less spillover.

Memory bandwidth is roughly ~8 TB/s per GPU, for about ~64 TB/s aggregate across the node. That combination — ~2.3 TB of high-bandwidth memory fed at ~64 TB/s — is what lets the HGX B300 keep more of a frontier-scale model resident and well-fed than either the B200 or the prior Hopper generation.

On compute, each B300 delivers ~4.5 PFLOPS of dense FP8 and ~13.5 PFLOPS of dense FP4 tensor throughput — about 36 and 108 PFLOPS respectively across the 8-GPU node, the figures NVIDIA’s DGX B300 datasheet states for a node. FP8 is level with the HGX B200; the gain is the FP4 path, a 1.5× step over the B200’s ~9 PFLOPS per GPU, aimed at high-throughput inference. All figures here are dense; the datasheet also quotes node totals "with sparsity," i.e. roughly 2× these. (The rack-scale GB300 NVL72 superchip runs a higher clock bin — roughly ~5 PFLOPS dense FP8 and ~15 PFLOPS dense FP4 per GPU — inside a 72-GPU NVLink domain rather than this 8-GPU node. Same silicon, different package and power envelope; do not read a rack figure onto a baseboard.)

Inside the node, the eight B300 GPUs are wired over a fifth-generation NVLink baseboard at 1.8 TB/s per GPU — the same per-GPU NVLink bandwidth as the HGX B200 — giving all-to-all connectivity across the 8-GPU NVLink domain with minimal communication overhead. This is a node-scale domain: unlike the GB300 NVL72 rack, whose 72 GPUs share one NVLink domain, the HGX B300 ties eight GPUs together per node.

To scale beyond a single node, the HGX B300 pairs with NVIDIA ConnectX-8 SuperNICs at 800 Gb/s per GPU with GPUDirect — double the per-GPU scale-out bandwidth of the B200’s ConnectX-7 (400 Gb/s) — for roughly ~6.4 Tb/s of scale-out bandwidth per node on InfiniBand or Ethernet fabrics. NVLink handles the dense traffic inside the node; ConnectX stitches many nodes into a larger training cluster.

Each B300 GPU on an HGX baseboard carries a board power of roughly ~1,100 W, a modest step over the ~1,000 W of the HGX B200. Treat that as a derived figure, not a published one: NVIDIA states no per-GPU TDP for the HGX or DGX B300 — the DGX B300 page gives only about 14 kW for the whole system — and ~1,100 W per GPU is what closes that budget. Eight GPUs at ~1,100 W is ~8.8 kW of accelerator, leaving roughly 5 kW of the system figure for two host CPUs, multiple terabytes of DDR5, the 800G NICs, the NVLink switch trays, fans, and power-supply losses.

Watch the bin. Blackwell Ultra ships in two power and clock bins, and which one a number belongs to is set by the package, not the chip. The 8-GPU HGX/DGX baseboard runs the ~1,100 W bin; the rack-scale GB300 NVL72 superchip runs a higher one at roughly ~1,400 W per GPU across 72 GPUs. The 1,400 W figure is widely quoted for "B300" and is correct only for the NVL72 rack — applied to an 8-GPU node it overstates the accelerator load by 2.4 kW and does not fit inside NVIDIA’s own system number.

So the air-versus-liquid choice here is not forced by a per-GPU thermal cliff — 1,100 W is well within what a cold plate or a well-designed air path can carry. It is a rack-and-facility decision, made at the node: ~14 kW in a single chassis is above what much of the installed air-cooled base was built for, so the constraint is the hall, not the GPU. That is why the HGX B300 ships both ways, and why the builds listed below include air- and liquid-cooled chassis side by side:

• Air-cooled. Drops into standard high-density racks with no facility water and no CDU — the fastest facilities fit, at lower rack density. The right answer whenever a site has no building loop.
• Liquid-cooled. Direct-to-chip packs the same eight GPUs into a shorter chassis and a denser rack, at the cost of a cooling loop or CDU. Chosen for density and for halls already plumbed for it.

Both carry the identical GPU, memory, and NVLink specs — only rack density and facility requirements differ. Exact per-node power draw depends on the integrator build, host CPU, and cooling, and is confirmed at quote.

The HGX B300 is the Blackwell Ultra step-up over the original Blackwell HGX B200 — same 8 GPUs per node and the same fifth-generation NVLink, with the gains concentrated in memory, scale-out, and inference throughput. See the full HGX B300 vs B200 comparison for when the step-up is worth it.

The HGX B300 ships as a factory-integrated 8-GPU node from several OEMs — Supermicro, Dell (PowerEdge), HPE (ProLiant XD690), and Lenovo (ThinkSystem SR680a V4) — 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 B300 build is quoted per configuration on request — the figure depends on integrator, cooling (air vs liquid), 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 B300 stacks up against the HGX B200 and the rack-scale GB300 NVL72.