NVIDIA HGX H200 Specs & Datasheet (8-GPU Node)

The HGX H200 is NVIDIA’s 8-GPU Hopper server platform — the memory-upgraded successor to the HGX H100, built for large-model training and, especially, memory-bound inference. Eight H200 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 H200 shares the Hopper compute architecture of the H100 but swaps in HBM3e memory: 141 GB per GPU instead of the H100’s 80 GB of HBM3, at higher bandwidth. That makes the HGX H200 a node you rack in a standard air-cooled data center — a 5U–6U chassis that drops into a conventional rack — and the mainstream Hopper platform for inference workloads whose bottleneck is memory capacity and bandwidth rather than raw compute. It is the Hopper predecessor to the Blackwell HGX B200 8-GPU node.

The HGX H200 8-GPU node specification, aggregated from the current integrator builds (HPE, 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 H200 GPU in the HGX H200 carries 141 GB of HBM3e — the on-package GPU memory (VRAM) available to the model. Across the 8-GPU baseboard that is 1,128 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 ~4.8 TB/s per GPU, and it is the headline reason to choose H200 over H100: the H200 keeps the same Hopper compute but nearly doubles memory capacity (141 GB vs 80 GB) and lifts bandwidth by moving from HBM3 to HBM3e. For memory-bound inference — where throughput is gated by how fast weights and KV cache stream out of VRAM — that extra capacity and bandwidth is what lets a single HGX H200 node serve larger models, longer contexts, and more concurrent requests than the H100 it replaces.

Compute is unchanged from the H100 — about 1,979 TFLOPS (~2 PFLOPS) of dense FP8 per GPU (~15.8 PFLOPS across the node) — since the H200 is a memory upgrade of the same Hopper die; the gain is entirely capacity and bandwidth. Sparsity doubles the FP8 figure.

Inside the node, the eight H200 GPUs are wired over a fourth-generation NVLink baseboard at 900 GB/s per GPU, giving all-to-all connectivity across the 8-GPU domain with minimal communication overhead — the same NVLink generation and bandwidth as the HGX H100.

To scale beyond a single node, the HGX H200 uses NVIDIA ConnectX-7 NDR400 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 fabrics. NVLink handles the dense traffic inside the node; ConnectX-7 NDR400 stitches many nodes into a larger training or inference cluster.

The HGX H200 ships as a factory-integrated, air-cooled 8-GPU node from several OEMs — HPE (Cray XD670), Dell (PowerEdge XE9680), Supermicro, and Gigabyte — each with its own chassis, host CPU, and support model. All are new and factory-integrated: the GPU, memory, and NVLink specifications are common across builds, while chassis, host CPU, GPU:NIC ratio, warranty, and lead time differ by integrator. Each build below links to its full specification.

Every HGX H200 build is quoted per configuration on request — the figure depends on integrator, 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 H200 compares to the HGX H100, or step up to the Blackwell HGX B200.