H100 or H200: memory for large models

H200 uses the same Hopper architecture with much more memory and higher bandwidth. The main advantage is that models and long contexts fit on fewer cards.

Specifications

	H100 SXM 80 GB	H200 NVL 141 GB
Video memory	80 GB	141 GB
Memory type	HBM3	HBM3e
Architecture	Hopper	Hopper
Price from	306₽/h	619₽/h
Servers now	21	23

Prices and availability come from the live catalog and update automatically.

Which one to pick

Pick H100 SXM if

budget matters more than memory
models up to 30–40B at fp16, or larger at int8
short context

Pick H200 NVL if

models that do not fit in 80 GB
long context and a large KV cache
fewer cards per instance

Cost example

H100 SXM, per hour: about 306 ₽
H200 NVL, per hour: about 619 ₽

Estimate at the current per-GPU hourly price. Billed per minute, the price is fixed at launch.

Questions and answers

Do I need an H200 for Llama 70B?: At fp16 the 70B weights take about 140 GB, so one H200 is not enough with headroom. At int8 or int4 it fits, and long context is easier to hold.
What is the difference between NVL and SXM?: The form factor and how cards are linked. Multi-GPU training depends on that link; single-card inference depends on it less.

H100 or H200: memory for large models

Specifications

Which one to pick

Pick H100 SXM if

Pick H200 NVL if

Cost example

Questions and answers

Compare with others

H100 or H200: memory for large models

Specifications

Which one to pick

Pick H100 SXM if

Pick H200 NVL if

Cost example

Questions and answers

Compare with others