← AI Hashrate

Best GPUs for 70B-class local LLMs

We treat '70B-class' as open-weight models with over 45B parameters — 24 in our catalog. The table ranks every catalog GPU that fully fits at least one of them at Q4 / 8K context, by average estimated decode tok/s across the models it fits. Fit means weights + KV cache + 1 GB runtime overhead within 95% of VRAM — not weights alone. Methodology.

#GPUVRAM (GB)Bandwidth (GB/s)MSRPFitsAvg Q4 tok/sBuy
1B200 192GB SXM1928,000$40,00013/24562.3 incl. measured
2MI300X 192GB1925,300$15,00013/24349.7 est.
3H200 141GB SXM51414,800$35,00011/24349.1 est.
4H100 80GB SXM5803,350$30,0009/24255.9 est.
5Gaudi 3 128GB1283,276$12,50010/24241.2 est.
6A100 80GB SXM4802,039$15,0009/24154.5 incl. measured
7RTX 6000 Ada 48GB48960$6,8002/24125.5 est.Search Amazon (affiliate)
8L40S 48GB48864$8,0002/24113.0 est.Search Amazon (affiliate)
9RTX A6000 48GB48768$4,5002/24100.4 est.Search Amazon (affiliate)
10RTX 5090 32GB321,792$1,9991/2488.4 est.Search Amazon (affiliate)
11Mac Studio M3 Ultra 96GB96819$3,99910/2461.0 incl. measuredSearch Amazon (affiliate)
12Mac Studio M2 Ultra 192GB192800$6,99913/2451.3 incl. measuredSearch Amazon (affiliate)
13MacBook Pro M4 Max 128GB128546$4,39910/2442.0 incl. measuredSearch Amazon (affiliate)
14MacBook Pro M3 Max 128GB128400$4,19910/2430.0 incl. measuredSearch Amazon (affiliate)
15Mac mini M4 Pro 64GB64273$2,1997/2420.2 incl. measuredSearch Amazon (affiliate)
16Mac mini M4 32GB32120$9991/245.9 est.Search Amazon (affiliate)

Averages mix estimated rows with measured anchors where they exist — per-model labels live on each hardware page. MSRP is launch list price, not live retail — Amazon prices change; click a Buy link for the current price. GPUs that fit none of this class are omitted.

More lists: 7B-class models · 13B-class models · 30B-class models · Under $500 · Under $1,000 · Under $2,000 · 8 GB VRAM · 12 GB VRAM · 16 GB VRAM · 24 GB VRAM · For Qwen3 · For DeepSeek · For Llama · For Gemma · For GLM · For coding · For reasoning · MoE picks