Llama on AMD Strix Halo

Can I run Llama 3.1 8B on an Ryzen AI Max+ 395 128GB?

Yes. Llama 3.1 8B at Q4_K_M uses 6.1 GB of the 96.0 GB available on a Ryzen AI Max+ 395 128GB, and runs at about 26.2 tokens per second. You can push the context to 128K.

Runs comfortably

6.1 GB of 96.0 GB · 6%
096 GB
Weights 4.6 GB
KV cache 1.0 GB
Runtime overhead 0.5 GB

Llama 3.1 8B at Q4_K_M leaves 89.9 GB spare on a Ryzen AI Max+ 395 128GB. There is room to raise the context length or move up a quantisation level.

Generation26.2tok/s
Prompt processing1308tok/s
Max context128Ktokens
KV per 1K tokens0GB

Every quantisation of Llama 3.1 8B on a Ryzen AI Max+ 395 128GB

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

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 15.0 GB 16.5 GB Runs comfortably 128K 8.68
INT8 / W8A8 8.50 7.7 GB 9.2 GB Runs comfortably 128K 16.3
Q8_0 (GGUF) 8.50 7.7 GB 9.2 GB Runs comfortably 128K 16.3
FP8 (E4M3) 8.00 7.5 GB 9.0 GB Runs comfortably 128K 16.8
Q6_K 6.56 6.1 GB 7.7 GB Runs comfortably 128K 20.2
AWQ 4-bit 4.25 5.4 GB 6.9 GB Runs comfortably 128K 22.7
GPTQ 4-bit 4.25 5.4 GB 6.9 GB Runs comfortably 128K 22.7
MXFP4 4.25 5.4 GB 6.9 GB Runs comfortably 128K 22.7
Q5_K_M 5.67 5.3 GB 6.8 GB Runs comfortably 128K 23.1
Q5_K_S 5.52 5.2 GB 6.7 GB Runs comfortably 128K 23.6
Q4_K_M 4.85 4.6 GB 6.1 GB Runs comfortably 128K 26.2
Q4_K_S 4.58 4.4 GB 5.9 GB Runs comfortably 128K 27.3
Q4_0 4.55 4.4 GB 5.9 GB Runs comfortably 128K 27.5
IQ4_XS 4.25 4.1 GB 5.6 GB Runs comfortably 128K 28.9
Q3_K_M 3.91 3.8 GB 5.4 GB Runs comfortably 128K 30.7
IQ3_M 3.70 3.7 GB 5.2 GB Runs comfortably 128K 32.0
IQ3_XXS 3.06 3.2 GB 4.7 GB Runs comfortably 128K 36.5
Q2_K 2.63 2.8 GB 4.3 GB Runs comfortably 128K 40.3
IQ2_XXS 2.06 2.3 GB 3.9 GB Runs comfortably 128K 46.8
IQ1_M 1.75 2.1 GB 3.6 GB Runs comfortably 128K 51.3

Also worth checking

Llama 3.1 8B on every GPU
Full hardware tableEverything that runs on a Ryzen AI Max+ 395 128GB
Full model tableWhy the KV cache is the part that bites
Guide