Llama · 70.5B parameters

Llama 3.1 70B VRAM requirements

Llama 3.1 70B has 80 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 40.0 GB.

Model

Hardware

Quantisation

Context 8K

KV cache precision

Runtime

Batch

GPUs

Monitor plugged into this GPU

Won't fit

43.4 GB of 21.8 GB · 199%

045 GB

Weights 40.0 GB

KV cache 2.5 GB

Runtime overhead 0.9 GB

Over the limit 21.6 GB

Short by 21.6 GB. You can run it with 36 of 80 layers on the RTX 4090 and the rest in system RAM, at roughly 1.60 tok/s — usable for batch work, painful for chat. A smaller quantisation or a shorter context is usually the better trade.

Generation1.60tok/s

Prompt processing491tok/s

Max context0tokens

KV per 1K tokens0GB

Every quantisation of Llama 3.1 70B 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	131.4 GB	134.8 GB	Won't fit	—	0.33
INT8 / W8A8	8.50	69.3 GB	72.7 GB	Won't fit	—	0.72
Q8_0 (GGUF)	8.50	69.3 GB	72.7 GB	Won't fit	—	0.72
FP8 (E4M3)	8.00	65.7 GB	69.1 GB	Won't fit	—	0.77
Q6_K	6.56	53.9 GB	57.3 GB	Won't fit	—	1.01
Q5_K_M	5.67	46.6 GB	50.0 GB	Won't fit	—	1.26
Q5_K_S	5.52	45.3 GB	48.7 GB	Won't fit	—	1.31
Q4_K_M	4.85	40.0 GB	43.4 GB	Won't fit	—	1.60
Q4_K_S	4.58	37.8 GB	41.2 GB	Won't fit	—	1.76
AWQ 4-bit	4.25	37.8 GB	41.2 GB	Won't fit	—	1.77
GPTQ 4-bit	4.25	37.8 GB	41.2 GB	Won't fit	—	1.77
MXFP4	4.25	37.8 GB	41.2 GB	Won't fit	—	1.77
Q4_0	4.55	37.6 GB	41.0 GB	Won't fit	—	1.81
IQ4_XS	4.25	35.2 GB	38.6 GB	Won't fit	—	2.02
Q3_K_M	3.91	32.5 GB	35.9 GB	Won't fit	—	2.39
IQ3_M	3.70	30.8 GB	34.2 GB	Won't fit	—	2.65
IQ3_XXS	3.06	25.7 GB	29.1 GB	Won't fit	—	4.26
Q2_K	2.63	22.3 GB	25.7 GB	Won't fit	—	6.78
IQ2_XXS	2.06	17.8 GB	21.2 GB	Fits, but tight	10K	31.3
IQ1_M	1.75	15.3 GB	18.7 GB	Runs comfortably	18K	35.9

Llama 3.1 70B 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	107K	55.4
A100 80GB	80	2039	Runs comfortably	107K	30.5
L40S	48	864	Fits, but tight	11K	12.5
RTX A6000	48	768	Fits, but tight	11K	11.6
Mac Studio M3 Ultra 256GB	256	819	Runs comfortably	128K	11.4
Mac Studio M4 Max 128GB	128	546	Runs comfortably	128K	8.48
NVIDIA DGX Spark (GB10)	128	273	Runs comfortably	128K	4.37
Ryzen AI Max+ 395 128GB	128	256	Runs comfortably	128K	3.26
RTX 5090	32	1792	Won't fit	—	2.65
Mac Mini M4 Pro 48GB	48	273	Won't fit	—	2.63
RTX 3090	24	936	Won't fit	—	1.67
RTX 4090	24	1008	Won't fit	—	1.60
Radeon RX 7900 XTX	24	960	Won't fit	—	1.52
RTX 5080	16	960	Won't fit	—	1.35
RTX 5070 Ti	16	896	Won't fit	—	1.35
RTX 5060 Ti 16GB	16	448	Won't fit	—	1.32
RTX 4080 Super	16	736	Won't fit	—	1.22
RTX 4070 Ti Super	16	672	Won't fit	—	1.22
RTX 5070	12	672	Won't fit	—	1.21
RTX 4060 Ti 16GB	16	288	Won't fit	—	1.17
RTX 3060 12GB	12	360	Won't fit	—	1.13
RTX 4070 Super	12	504	Won't fit	—	1.09
RTX 4070	12	504	Won't fit	—	1.09
RTX 3080 10GB	10	760	Won't fit	—	1.09
Arc B580	12	456	Won't fit	—	0.92

Architecture

Parameters	70.5B
Layers	80
Hidden size	8192
Attention heads / KV heads	64 / 8
Head dimension	128
Vocabulary	128,256
Trained context	128K
KV cache per 1K tokens	0 GB
Hugging Face	meta-llama/Llama-3.1-70B-Instruct

Direct answers

Llama 3.1 70B on RTX 5090
See the verdict Llama 3.1 70B on RTX 4090
See the verdict Llama 3.1 70B on RTX 3090
See the verdict Llama 3.1 70B on RTX 5080
See the verdict Llama 3.1 70B on RTX 5070 Ti
See the verdict Llama 3.1 70B on RTX 5070
See the verdict Llama 3.1 70B on RTX 4070 Ti Super
See the verdict Llama 3.1 70B on RTX 4070 Super
See the verdict Llama 3.1 70B on RTX 4070
See the verdict