Same Hopper silicon, same 700W, same compute — the H200 is an H100 with much more, faster memory. This is a memory decision: 80GB HBM3 at 3.4 TB/s versus 141GB HBM3e at 4.8 TB/s. Here is exactly when the extra memory is worth it.
Every spec side by side. Hard specs mirror the live catalog; confirm exact configuration at quote.
| Specification | H100 | H200 |
|---|---|---|
| Architecture | Hopper (GH100, TSMC 4N) | Hopper (GH100, TSMC 4N) |
| GPUs per unit | 1 (8 per HGX/DGX server) | 1 (8 per HGX/DGX server) |
| Memory | 80 GB HBM3 | 141 GB HBM3e |
| Memory bandwidth | 3.4 TB/s | 4.8 TB/s |
| FP8 tensor | 3,958 TFLOPS (with sparsity) | 3,958 TFLOPS (with sparsity) |
| FP16 / BF16 tensor | 1,979 TFLOPS (with sparsity) | 1,979 TFLOPS (with sparsity) |
| NVLink | 900 GB/s (4th-gen NVLink) | 900 GB/s (4th-gen NVLink) |
| Board power (TDP) | 700 W | 700 W |
| Form factor | SXM5 | SXM5 |
| Cooling | Air or direct-to-chip liquid | Air or direct-to-chip liquid |
| Typical lead time | 10–14 wk | 10–14 wk |
| Indicative price | From $26,500 / unit | From $31,000 / unit |
The decision comes down to workload, facility and timeline. Here is the call by scenario.
Larger models, bigger KV caches and longer contexts fit per GPU — the 141GB / 4.8 TB/s memory lifts real-world tokens-per-second and can cut the GPU count for a serving target.
Identical training throughput to the H200 at a lower price, with a deep new/used/refurbished market and shorter leads. The pragmatic pick when memory is not the bottleneck.
Both are 700W and share the SXM5 socket, NVLink and cooling design — mix them in the same infrastructure and upgrade memory where it pays.
Open either product, add it to your quote list — free, no commitment — or size the facility with our calculators first.
If your workload is memory-bound — inference, long context, or models that barely fit — yes: 141GB HBM3e at 4.8 TB/s meaningfully raises throughput and can reduce GPU count. For compute-bound training the H100 delivers the same TFLOPS for less money.
They have identical compute (same GH100 die, same FP8/FP16 TFLOPS). The H200 only differs in memory: 141GB HBM3e at 4.8 TB/s versus 80GB HBM3 at 3.4 TB/s, which matters on memory-bound work.
Yes. Both are 700W SXM5 with 4th-gen NVLink and share server, cooling and power designs, so they coexist in the same infrastructure — a common way to add memory headroom without re-architecting a hall.
Our solutions engineers size compute, power and cooling together and confirm real lead times — no payment, no commitment, quotes back in about one business day.
Our engineers scope power, cooling and compute together. No payment, and quotes come back in about a business day.