High-end tier · NVIDIA H200 141GB
NVIDIA H200 141GB
Hopper compute, lifted by 76% more memory. H200's 141 GB of HBM3e and 4.8 TB/s of bandwidth let serious training clusters run 70B-class models with headroom — without stepping up to Blackwell economics.
ArchitectureHopper
Memory141 GB HBM3e
Bandwidth4.8 TB/s
FP16 dense989 TFLOPS
Specifications
H200, by the numbers.
| Spec | NVIDIA H200 (Hopper) |
|---|---|
| Architecture | NVIDIA Hopper (GH100) |
| Memory | 141 GB HBM3e |
| Memory bandwidth | 4.8 TB/s |
| FP8 (dense) | 1,979 TFLOPS |
| FP16 / BF16 (dense) | 989 TFLOPS |
| FP64 (dense) | 34 TFLOPS |
| NVLink | 900 GB/s per GPU |
| Form factor | 8-GPU HGX baseboard, 4U node |
| Power envelope | 700 W per accelerator |
| Boot & security | TPM 2.0, verified boot, encrypted RAM |
Use cases
Where H200 shines.
70B-class training
FSDP and tensor-parallel training for large open models, scaled out over NVLink-connected fabrics.
Long-context workloads
Big memory for long sequences, huge batches, and dense retrieval pipelines that spill over smaller cards.
Inference with headroom
Serve large models with long context windows — 141 GB holds weights and KV cache without sharding.
Ready to push the frontier?
Provision bare-metal GPUs in minutes, scale capacity on demand, and get a real engineer on call — not a ticket queue.