Llama on NVIDIA Ampere

Can I run Llama 3.1 8B on an RTX 3060 12GB?

Yes. Llama 3.1 8B at Q4_K_M uses 6.4 GB of the 10.5 GB available on a RTX 3060 12GB, and runs at about 43.7 tokens per second. You can push the context to 40K.

Runs comfortably

6.4 GB of 10.5 GB · 61%
010 GB
Weights 4.6 GB
KV cache 1.0 GB
Runtime overhead 0.8 GB

Llama 3.1 8B at Q4_K_M leaves 4.0 GB spare on a RTX 3060 12GB. There is room to raise the context length or move up a quantisation level.

Generation43.7tok/s
Prompt processing654tok/s
Max context40Ktokens
KV per 1K tokens0GB

Every quantisation of Llama 3.1 8B on a RTX 3060 12GB

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

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 15.0 GB 16.8 GB Won't fit — 4.81
INT8 / W8A8 8.50 7.7 GB 9.5 GB Fits, but tight 16K 27.3
Q8_0 (GGUF) 8.50 7.7 GB 9.5 GB Fits, but tight 16K 27.3
FP8 (E4M3) 8.00 7.5 GB 9.3 GB Runs comfortably 17K 28.1
Q6_K 6.56 6.1 GB 8.0 GB Runs comfortably 28K 33.8
AWQ 4-bit 4.25 5.4 GB 7.2 GB Runs comfortably 34K 37.9
GPTQ 4-bit 4.25 5.4 GB 7.2 GB Runs comfortably 34K 37.9
MXFP4 4.25 5.4 GB 7.2 GB Runs comfortably 34K 37.9
Q5_K_M 5.67 5.3 GB 7.1 GB Runs comfortably 35K 38.6
Q5_K_S 5.52 5.2 GB 7.0 GB Runs comfortably 36K 39.5
Q4_K_M 4.85 4.6 GB 6.4 GB Runs comfortably 40K 43.7
Q4_K_S 4.58 4.4 GB 6.2 GB Runs comfortably 42K 45.6
Q4_0 4.55 4.4 GB 6.2 GB Runs comfortably 42K 45.9
IQ4_XS 4.25 4.1 GB 5.9 GB Runs comfortably 44K 48.3
Q3_K_M 3.91 3.8 GB 5.7 GB Runs comfortably 46K 51.3
IQ3_M 3.70 3.7 GB 5.5 GB Runs comfortably 48K 53.4
IQ3_XXS 3.06 3.2 GB 5.0 GB Runs comfortably 52K 60.9
Q2_K 2.63 2.8 GB 4.6 GB Runs comfortably 55K 67.2
IQ2_XXS 2.06 2.3 GB 4.2 GB Runs comfortably 59K 78.0
IQ1_M 1.75 2.1 GB 3.9 GB Runs comfortably 61K 85.4

Also worth checking

Llama 3.1 8B on every GPU
Full hardware tableEverything that runs on a RTX 3060 12GB
Full model tableWhy the KV cache is the part that bites
Guide