NVIDIA HGX H100 Specs & Datasheet (8-GPU Node)

The HGX H100 is NVIDIA’s 8-GPU Hopper server platform — the workhorse data-center node that anchored the current generative-AI training and inference buildout. Eight H100 SXM GPUs sit on a single fourth-generation NVLink baseboard, so the node behaves as one tightly-coupled 8-GPU accelerator rather than eight discrete cards.

The HGX H100 is a node you rack in a standard data center: a chassis that drops into a conventional rack, offered by multiple OEMs in both air- and liquid-cooled builds. It is the Hopper predecessor to the memory-upgraded HGX H200 (same Hopper compute, HBM3e instead of HBM3) and to the Blackwell HGX B200 8-GPU node. Each H100 SXM module carries 80 GB of HBM3 and draws 700 W.

The HGX H100 8-GPU node specification, aggregated from the current integrator builds (Supermicro, Dell, Gigabyte, HPE). 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. Per-node compute and memory totals are the per-GPU figures multiplied across the eight GPUs.

Each H100 GPU in the HGX H100 carries 80 GB of HBM3 — the on-package GPU memory (VRAM) available to the model. Across the 8-GPU baseboard that is 640 GB of HBM3 per node, pooled and addressable over NVLink so a large model and its KV cache can span all eight GPUs.

Memory bandwidth is 3.35 TB/s per GPU — the rate at which weights and activations stream out of HBM3. For memory-bound inference, where throughput is gated by how fast the model streams out of VRAM, that bandwidth (together with the 80 GB capacity) is the ceiling on how large a model and how long a context a single H100 can serve. The HGX H200 keeps the identical Hopper compute but lifts each GPU to 141 GB of HBM3e at higher bandwidth — the memory-focused reason to step up from H100 to H200.

Inside the node, the eight H100 GPUs are wired over a fourth-generation NVLink baseboard at 900 GB/s per GPU, giving all-to-all connectivity across the 8-GPU NVLink domain with minimal communication overhead. (This is the Hopper generation — fourth-gen NVLink at 900 GB/s and an 8-GPU domain — not the fifth-generation NVLink of the Blackwell HGX B200.)

To scale beyond a single node, the HGX H100 uses NVIDIA ConnectX-7 NDR400 adapters at 400 Gb/s each with GPUDirect — for roughly ~3.2 Tb/s of scale-out bandwidth per node on InfiniBand or Ethernet fabrics. NVLink handles the dense traffic inside the node; ConnectX-7 stitches many nodes into a larger training or inference cluster. Exact GPU:NIC ratio and fabric are confirmed at quote.

Each H100 SXM module has a 700 W TDP, so the eight GPUs alone account for roughly 5.6 kW before the host CPUs, memory, NICs, and fans — the HGX H100 is a high-density node, and the exact per-node power draw depends on the integrator build and host platform, confirmed at quote.

The platform ships in both air- and liquid-cooled builds: Supermicro offers the H100 SXM board in an air-cooled chassis and a direct-to-chip liquid-cooled variant, and Dell, Gigabyte, and HPE each ship their own thermal design. Both carry the identical GPU, memory, and NVLink specs — only density and facility requirements differ. See HGX H100 air vs liquid cooling for the trade-off: liquid buys density at the cost of a cooling loop, while air is simpler to deploy.

The HGX H100 and HGX H200 are the same Hopper compute — identical dense FP16 (~989.5 TFLOPS) and FP8 (1,979 TFLOPS) per GPU, the same fourth-generation NVLink at 900 GB/s, and the same 700 W SXM module. The difference is memory: the H100 has 80 GB of HBM3 at 3.35 TB/s per GPU, while the H200 has 141 GB of HBM3e at ~4.8 TB/s. The H200 is a memory upgrade of the same die — choose it when the workload is memory-bound (larger models, longer contexts, more concurrent requests); the H100 remains the mainstream Hopper node for compute-bound training and inference. See the full HGX H200 vs H100 comparison.

The HGX H100 ships as a factory-integrated 8-GPU node from several OEMs — Supermicro (air and liquid), Dell, Gigabyte, and HPE — each with its own chassis, cooling, host CPU, and support model. The GPU, memory, and NVLink specifications are common across builds, while chassis, cooling, host CPU, GPU:NIC ratio, warranty, and lead time differ by integrator. Each build below links to its full specification.

Every HGX H100 build is quoted per configuration on request — the figure depends on integrator, cooling (air vs liquid), host CPU, GPU:NIC ratio, 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, see how the H100 compares to the HGX H200, or read how Hopper stacks up against Blackwell.