FR
← All hardware
Datacenter · HBM3

H100 80 GB: which LLMs can it run?

On its own, it loads up to gpt-oss-120b in Q4. With several cards, memory adds up and opens bigger models.

Specifications

Memory80 GB
Bandwidth3,350 GB/s
FP16 compute~989 TF
Power700 W
Indicative price€28,000
Cloud rental€2.80/h
Build a server with it →Compare

Against cards at the same price

H100 80 GB vs H200 141 GB€33,000
H100 80 GB vs B200 192 GB€42,000
H100 80 GB vs Mac Studio M3 Ultra 512 GB€11,500

Models and expected speed

ModelQ4Q8FP16
Qwen3 4B✓ ~723 tok/s✓ ~420 tok/s✓ ~226 tok/s
Llama 3.1 8B✓ ~372 tok/s✓ ~213 tok/s✓ ~114 tok/s
Qwen3 8B✓ ~363 tok/s✓ ~208 tok/s✓ ~111 tok/s
Gemma 3 12B✓ ~251 tok/s✓ ~143 tok/s✓ ~76 tok/s
Mistral Nemo 12B✓ ~247 tok/s✓ ~141 tok/s✓ ~75 tok/s
Phi-4 14B✓ ~205 tok/s✓ ~117 tok/s✓ ~62 tok/s
Qwen3 14B✓ ~204 tok/s✓ ~116 tok/s✓ ~62 tok/s
gpt-oss-20b✓ ~798 tok/s✓ ~465 tok/s✓ ~251 tok/s
Mistral Small 24B✓ ~127 tok/s✓ ~72 tok/s✓ ~38 tok/s
Gemma 3 27B✓ ~113 tok/s✓ ~64 tok/s✓ ~34 tok/s
Qwen3.8 27B✓ ~109 tok/s✓ ~62 tok/s✓ ~33 tok/s
Qwen3 30B-A3B✓ ~865 tok/s✓ ~505 tok/s✓ ~273 tok/s
Qwen3 32B✓ ~93 tok/s✓ ~53 tok/s✓ ~28 tok/s
Qwen2.5 Coder 32B✓ ~93 tok/s✓ ~53 tok/s✓ ~28 tok/s
DeepSeek R1 Distill 32B✓ ~93 tok/s✓ ~53 tok/s✓ ~28 tok/s
Mixtral 8x7B✓ ~234 tok/s✓ ~133 tok/s2 × · ~107 tok/s
Llama 3.3 70B✓ ~43 tok/s2 × · ~37 tok/s2 × · ~20 tok/s
GLM-4.5-Air 106B✓ ~251 tok/s2 × · ~215 tok/s3 × · ~172 tok/s
gpt-oss-120b✓ ~574 tok/s2 × · ~497 tok/s4 × · ~534 tok/s
Mistral Large 123B2 × · ~37 tok/s2 × · ~21 tok/s4 × · ~22 tok/s
Qwen3 235B-A22B2 × · ~207 tok/s4 × · ~236 tok/s7 × · ~219 tok/s
GLM-4.5 355B3 × · ~214 tok/s6 × · ~243 tok/s✗
Llama 3.1 405B4 × · ~23 tok/s6 × · ~19 tok/s✗
DeepSeek R1 671B6 × · ~371 tok/s✗✗
DeepSeek V3.1 671B6 × · ~371 tok/s✗✗
Kimi K2 1T✗✗✗

8k tokens, one user. “2 ×” = number of cards needed, speed with tensor parallelism (vLLM). ✗ = doesn’t fit, even with 8 cards.