Llama on NVIDIA Hopper

Can I run Llama 3.1 70B on an H100 SXM 80GB?

Yes. Llama 3.1 70B at Q4_K_M uses 43.4 GB of the 74.4 GB available on a H100 SXM 80GB, and runs at about 55.4 tokens per second. You can push the context to 107K.

Runs comfortably

43.4 GB of 74.4 GB · 58%
074 GB
Weights 40.0 GB
KV cache 2.5 GB
Runtime overhead 0.9 GB

Llama 3.1 70B at Q4_K_M leaves 31.0 GB spare on a H100 SXM 80GB. There is room to raise the context length or move up a quantisation level.

Generation55.4tok/s
Prompt processing2944tok/s
Max context107Ktokens
KV per 1K tokens0GB

Every quantisation of Llama 3.1 70B on a H100 SXM 80GB

Highlighted row is the highest quality that still fits at 8K context.

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 131.4 GB 134.8 GB Won't fit — 0.71
INT8 / W8A8 8.50 69.3 GB 72.7 GB Fits, but tight 13K 32.7
Q8_0 (GGUF) 8.50 69.3 GB 72.7 GB Fits, but tight 13K 32.7
FP8 (E4M3) 8.00 65.7 GB 69.1 GB Fits, but tight 25K 34.5
Q6_K 6.56 53.9 GB 57.3 GB Runs comfortably 63K 41.7
Q5_K_M 5.67 46.6 GB 50.0 GB Runs comfortably 86K 47.9
Q5_K_S 5.52 45.3 GB 48.7 GB Runs comfortably 90K 49.2
Q4_K_M 4.85 40.0 GB 43.4 GB Runs comfortably 107K 55.4
Q4_K_S 4.58 37.8 GB 41.2 GB Runs comfortably 114K 58.3
AWQ 4-bit 4.25 37.8 GB 41.2 GB Runs comfortably 114K 58.4
GPTQ 4-bit 4.25 37.8 GB 41.2 GB Runs comfortably 114K 58.4
MXFP4 4.25 37.8 GB 41.2 GB Runs comfortably 114K 58.4
Q4_0 4.55 37.6 GB 41.0 GB Runs comfortably 115K 58.7
IQ4_XS 4.25 35.2 GB 38.6 GB Runs comfortably 123K 62.4
Q3_K_M 3.91 32.5 GB 35.9 GB Runs comfortably 128K 67.3
IQ3_M 3.70 30.8 GB 34.2 GB Runs comfortably 128K 70.7
IQ3_XXS 3.06 25.7 GB 29.1 GB Runs comfortably 128K 83.5
Q2_K 2.63 22.3 GB 25.7 GB Runs comfortably 128K 95.1
IQ2_XXS 2.06 17.8 GB 21.2 GB Runs comfortably 128K 117
IQ1_M 1.75 15.3 GB 18.7 GB Runs comfortably 128K 133

Also worth checking

Llama 3.1 70B on every GPU
Full hardware tableEverything that runs on a H100 SXM 80GB
Full model tableWhy the KV cache is the part that bites
Guide