FR
← All hardware
Datacenter · HBM2e

A100 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
Bandwidth2,039 GB/s
FP16 compute~312 TF
Power300 W
Indicative price€11,000
Cloud rental€1.60/h
Build a server with it →Compare

Against cards at the same price

A100 80 GB vs Mac Studio M3 Ultra 512 GB€11,500
A100 80 GB vs RTX PRO 6000 Blackwell 96 GB€9,000
A100 80 GB vs L40S 48 GB€7,500

Models and expected speed

ModelQ4Q8FP16
Qwen3 4B✓ ~440 tok/s✓ ~255 tok/s✓ ~138 tok/s
Llama 3.1 8B✓ ~227 tok/s✓ ~130 tok/s✓ ~69 tok/s
Qwen3 8B✓ ~221 tok/s✓ ~127 tok/s✓ ~68 tok/s
Gemma 3 12B✓ ~153 tok/s✓ ~87 tok/s✓ ~46 tok/s
Mistral Nemo 12B✓ ~150 tok/s✓ ~86 tok/s✓ ~46 tok/s
Phi-4 14B✓ ~125 tok/s✓ ~71 tok/s✓ ~38 tok/s
Qwen3 14B✓ ~124 tok/s✓ ~71 tok/s✓ ~38 tok/s
gpt-oss-20b✓ ~485 tok/s✓ ~283 tok/s✓ ~153 tok/s
Mistral Small 24B✓ ~77 tok/s✓ ~44 tok/s✓ ~23 tok/s
Gemma 3 27B✓ ~69 tok/s✓ ~39 tok/s✓ ~21 tok/s
Qwen3.8 27B✓ ~66 tok/s✓ ~38 tok/s✓ ~20 tok/s
Qwen3 30B-A3B✓ ~527 tok/s✓ ~307 tok/s✓ ~166 tok/s
Qwen3 32B✓ ~57 tok/s✓ ~32 tok/s✓ ~17 tok/s
Qwen2.5 Coder 32B✓ ~57 tok/s✓ ~32 tok/s✓ ~17 tok/s
DeepSeek R1 Distill 32B✓ ~57 tok/s✓ ~32 tok/s✓ ~17 tok/s
Mixtral 8x7B✓ ~142 tok/s✓ ~81 tok/s2 × · ~65 tok/s
Llama 3.3 70B✓ ~26 tok/s2 × · ~22 tok/s2 × · ~12 tok/s
GLM-4.5-Air 106B✓ ~153 tok/s2 × · ~131 tok/s3 × · ~104 tok/s
gpt-oss-120b✓ ~349 tok/s2 × · ~303 tok/s4 × · ~325 tok/s
Mistral Large 123B2 × · ~23 tok/s2 × · ~13 tok/s4 × · ~14 tok/s
Qwen3 235B-A22B2 × · ~126 tok/s4 × · ~143 tok/s7 × · ~133 tok/s
GLM-4.5 355B3 × · ~130 tok/s6 × · ~148 tok/s✗
Llama 3.1 405B4 × · ~14 tok/s6 × · ~12 tok/s✗
DeepSeek R1 671B6 × · ~226 tok/s✗✗
DeepSeek V3.1 671B6 × · ~226 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.