Compute · Large-memory Hopper

NVIDIA HGX H200

141 GB of HBM3e per GPU — serve and train large models without quantization compromises, on a mature Hopper platform.

Specifications

Memory first

GPUs per node8 × H200 (SXM)
GPU memory141 GB HBM3e per GPU
Memory bandwidth4.8 TB/s per GPU
Node interconnectNVLink 4, NVSwitch
Cluster interconnect400 Gb/s InfiniBand per GPU (Quantum-2)
AvailabilityOn-demand & reserved

Interconnect

400 Gb/s per GPU, non-blocking

Quantum-2 InfiniBand with GPUDirect RDMA keeps multi-node H200 clusters scaling linearly — the network never becomes the distributed-training bottleneck.

HGX H200 nodes in the data hall

Suited workloads

Large-context inference LLM training & fine-tuning Memory-bound models RAG at scale

Start your journey today

Tell us about your workload — an engineer, not a sales bot, will get back to you within one business day.