Llama · 8.0B parameters

Llama 3.1 8B VRAM requirements

Llama 3.1 8B has 32 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 4.6 GB, and the best quantisation that fits a 24 GB card is FP16 / BF16.

Model

Hardware

Quantisation

Context 8K

KV cache precision

Runtime

Batch

GPUs

Monitor plugged into this GPU

Runs comfortably

6.4 GB of 21.8 GB · 30%

022 GB

Weights 4.6 GB

KV cache 1.0 GB

Runtime overhead 0.8 GB

Llama 3.1 8B at Q4_K_M leaves 15.3 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.

Generation116tok/s

Prompt processing4315tok/s

Max context128Ktokens

KV per 1K tokens0GB

Every quantisation of Llama 3.1 8B on a RTX 4090

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

Quantisation	bpw	Weights	Total	Verdict	Max ctx	tok/s
FP16 / BF16	16.00	15.0 GB	16.8 GB	Runs comfortably	48K	38.8
INT8 / W8A8	8.50	7.7 GB	9.5 GB	Runs comfortably	106K	72.6
Q8_0 (GGUF)	8.50	7.7 GB	9.5 GB	Runs comfortably	106K	72.6
FP8 (E4M3)	8.00	7.5 GB	9.3 GB	Runs comfortably	108K	74.7
Q6_K	6.56	6.1 GB	8.0 GB	Runs comfortably	118K	89.6
AWQ 4-bit	4.25	5.4 GB	7.2 GB	Runs comfortably	124K	100
GPTQ 4-bit	4.25	5.4 GB	7.2 GB	Runs comfortably	124K	100
MXFP4	4.25	5.4 GB	7.2 GB	Runs comfortably	124K	100
Q5_K_M	5.67	5.3 GB	7.1 GB	Runs comfortably	125K	102
Q5_K_S	5.52	5.2 GB	7.0 GB	Runs comfortably	126K	105
Q4_K_M	4.85	4.6 GB	6.4 GB	Runs comfortably	128K	116
Q4_K_S	4.58	4.4 GB	6.2 GB	Runs comfortably	128K	121
Q4_0	4.55	4.4 GB	6.2 GB	Runs comfortably	128K	121
IQ4_XS	4.25	4.1 GB	5.9 GB	Runs comfortably	128K	127
Q3_K_M	3.91	3.8 GB	5.7 GB	Runs comfortably	128K	135
IQ3_M	3.70	3.7 GB	5.5 GB	Runs comfortably	128K	141
IQ3_XXS	3.06	3.2 GB	5.0 GB	Runs comfortably	128K	160
Q2_K	2.63	2.8 GB	4.6 GB	Runs comfortably	128K	176
IQ2_XXS	2.06	2.3 GB	4.2 GB	Runs comfortably	128K	204
IQ1_M	1.75	2.1 GB	3.9 GB	Runs comfortably	128K	223

Llama 3.1 8B on each GPU

Q4_K_M weights at 8K context, single card, monitor attached.

GPU	VRAM	GB/s	Verdict	Max ctx	tok/s
H100 SXM 80GB	80	3350	Runs comfortably	128K	422
A100 80GB	80	2039	Runs comfortably	128K	238
RTX 5090	32	1792	Runs comfortably	128K	222
RTX 5080	16	960	Runs comfortably	70K	121
RTX 4090	24	1008	Runs comfortably	128K	116
RTX 5070 Ti	16	896	Runs comfortably	70K	113
RTX 3090	24	936	Runs comfortably	128K	112
Radeon RX 7900 XTX	24	960	Runs comfortably	128K	105
L40S	48	864	Runs comfortably	128K	99.4
RTX A6000	48	768	Runs comfortably	128K	92.3
RTX 3080 10GB	10	760	Runs comfortably	25K	91.4
Mac Studio M3 Ultra 256GB	256	819	Runs comfortably	128K	90.9
RTX 5070	12	672	Runs comfortably	40K	85.4
RTX 4080 Super	16	736	Runs comfortably	70K	84.9
RTX 4070 Ti Super	16	672	Runs comfortably	70K	77.6
Mac Studio M4 Max 128GB	128	546	Runs comfortably	128K	67.8
RTX 4070 Super	12	504	Runs comfortably	40K	58.4
RTX 4070	12	504	Runs comfortably	40K	58.4
RTX 5060 Ti 16GB	16	448	Runs comfortably	70K	57.3
Arc B580	12	456	Runs comfortably	40K	44.9
RTX 3060 12GB	12	360	Runs comfortably	40K	43.7
NVIDIA DGX Spark (GB10)	128	273	Runs comfortably	128K	35.1
Mac Mini M4 Pro 48GB	48	273	Runs comfortably	128K	34.1
RTX 4060 Ti 16GB	16	288	Runs comfortably	70K	33.6
Ryzen AI Max+ 395 128GB	128	256	Runs comfortably	128K	26.2

Architecture

Parameters	8.0B
Layers	32
Hidden size	4096
Attention heads / KV heads	32 / 8
Head dimension	128
Vocabulary	128,256
Trained context	128K
KV cache per 1K tokens	0 GB
Hugging Face	meta-llama/Llama-3.1-8B-Instruct

Direct answers

Llama 3.1 8B on RTX 5090
See the verdict Llama 3.1 8B on RTX 4090
See the verdict Llama 3.1 8B on RTX 3090
See the verdict Llama 3.1 8B on RTX 5080
See the verdict Llama 3.1 8B on RTX 5070 Ti
See the verdict Llama 3.1 8B on RTX 5070
See the verdict Llama 3.1 8B on RTX 4070 Ti Super
See the verdict Llama 3.1 8B on RTX 4070 Super
See the verdict Llama 3.1 8B on RTX 4070
See the verdict