Llama-3.1-8B-Instruct — which GPUs fit

Meta · 8.0B · ctx 128K · llama3.1 · HuggingFace

Llama-3.1-8B-Instruct (Meta) is a 8.0B parameters open-weight model tracked on AI Hashrate. At 8K context we estimate roughly 7.0 GB VRAM for Q4 (weights 4.4 GB + KV 1.6 GB + 1 GB overhead) and 18.6 GB for FP16. About 50 GPUs/accelerators in our catalog fully fit this model at Q4 under that context assumption. Fastest Q4 configs: B200 192GB SXM ≈ 636.4 tok/s (estimated); MI300X 192GB ≈ 421.6 tok/s (estimated); H200 141GB SXM5 ≈ 381.8 tok/s (estimated); H100 80GB SXM5 ≈ 266.5 tok/s (estimated); Gaudi 3 128GB ≈ 260.6 tok/s (estimated). Use the table below for VRAM fit and tok/s per dollar (list MSRP). Estimates follow memory-bandwidth math; measured rows override when present. See methodology for details. Methodology.

100 hardware configs — sorted by tok/s. Default fit context 8K.

Hardware	VRAM	Quant	tok/s	Fits?	tok/s/$
B200 192GB SXM	192	Q4	636.4 est.	Yes	0.016
MI300X 192GB	192	Q4	421.6 est.	Yes	0.028
H200 141GB SXM5	141	Q4	381.8 est.	Yes	0.011
H100 80GB SXM5	80	FP16	350.0 measured	Yes	0.012
MI300X 192GB	192	FP16	320.0 measured	Yes	0.021
H100 80GB SXM5	80	Q4	266.5 est.	Yes	0.009
Gaudi 3 128GB	128	Q4	260.6 est.	Yes	0.021
RTX 5090 32GB	32	Q4	214.3 measured	Yes	0.107
A100 80GB SXM4	80	FP16	180.0 measured	Yes	0.012
B200 192GB SXM	192	FP16	175.0 est.	Yes	0.004
A100 80GB SXM4	80	Q4	162.2 est.	Yes	0.011
RTX 4090 24GB	24	Q4	145.0 measured	Yes	0.091
RTX 6000 Ada 48GB	48	Q4	138.6 measured	Yes	0.02
RTX 5080 16GB	16	Q4	136.2 measured	Yes	0.136
RTX 5070 Ti 16GB	16	Q4	131.9 measured	Yes	0.176
H200 141GB SXM5	141	FP16	105.0 est.	Yes	0.003
RTX 5070 12GB	12	Q4	100.4 measured	Yes	0.183
RTX 4070 Ti Super 16GB	16	Q4	95.0 measured	Yes	0.119
RTX 3090 24GB	24	Q4	92.0 measured	Yes	0.061
RX 9070 XT 16GB	16	Q4	89.1 measured	Yes	0.149
RX 9070 16GB	16	Q4	87.6 measured	Yes	0.16
RTX 4090D 24GB (China)	24	Q4	80.2 est.	Yes	0.05
RTX 3090 Ti 24GB	24	Q4	80.2 est.	Yes	0.04
RTX 5060 Ti 16GB	16	Q4	74.7 measured	Yes	0.174
Gaudi 3 128GB	128	FP16	71.7 est.	Yes	0.006
RTX 3080 10GB	10	Q4	70.0 measured	Yes	0.1
L40S 48GB	48	Q4	68.7 est.	Yes	0.009
RX 7800 XT 16GB	16	Q4	68.0 measured	Yes	0.136
Mac Studio M3 Ultra 96GB	96	Q4	65.1 est.	Yes	0.016
RX 7900 XTX 24GB	24	Q4	65.0 measured	Yes	0.065
Mac Studio M2 Ultra 192GB	192	Q4	63.6 est.	Yes	0.009
RX 7900 XT 20GB	20	Q4	63.6 est.	Yes	0.071
RTX A6000 48GB	48	Q4	61.1 est.	Yes	0.014
RTX 4080 Super 16GB	16	Q4	58.0 measured	Yes	0.058
RTX 4070 12GB	12	Q4	52.1 measured	Yes	0.087
RTX 4060 Ti 16GB	16	Q4	50.0 measured	Yes	0.1
RX 7900 GRE 16GB	16	Q4	45.8 est.	Yes	0.083
Arc A770 16GB	16	Q4	44.5 est.	Yes	0.128
RTX 5080 16GB	16	FP16	44.4 est.	No	0.044
RTX 4070 Ti 12GB	12	Q4	44.1 est.	Yes	0.055
MacBook Pro M4 Max 128GB	128	Q4	43.4 est.	Yes	0.01
RTX 3060 12GB	12	Q4	42.0 measured	Yes	0.128
RTX 5070 Ti 16GB	16	FP16	41.4 est.	No	0.055
RTX 4070 Super 12GB	12	Q4	40.1 est.	Yes	0.067
RTX 5090 32GB	32	FP16	39.2 est.	Yes	0.02
Arc B580 12GB	12	Q4	36.3 est.	Yes	0.146
RTX 5060 8GB	8	Q4	35.6 est.	Yes	0.119
RTX 5060 Ti 8GB	8	Q4	35.6 est.	Yes	0.094
RTX 3070 8GB	8	Q4	35.6 est.	Yes	0.071
MacBook Pro M3 Max 128GB	128	Q4	35.0 measured	Yes	0.008
RX 7700 XT 12GB	12	Q4	34.4 est.	Yes	0.077
RTX 4080 Super 16GB	16	FP16	34.0 est.	No	0.034
RTX 4070 Ti Super 16GB	16	FP16	31.1 est.	No	0.039
Arc B570 10GB	10	Q4	30.2 est.	Yes	0.138
RX 9070 16GB	16	FP16	29.6 est.	No	0.054
RX 9070 XT 16GB	16	FP16	29.6 est.	No	0.049
RX 7800 XT 16GB	16	FP16	28.9 est.	No	0.058
RX 7900 GRE 16GB	16	FP16	26.6 est.	No	0.048
Arc A770 16GB	16	FP16	25.9 est.	No	0.074
T4 16GB	16	Q4	25.5 est.	Yes	0.013
RTX 4060 Ti 8GB	8	Q4	22.9 est.	Yes	0.057
RX 7600 8GB	8	Q4	22.9 est.	Yes	0.085
RX 7600 XT 16GB	16	Q4	22.9 est.	Yes	0.07
RTX 4090 24GB	24	FP16	22.0 est.	Yes	0.014
RTX 4090D 24GB (China)	24	FP16	22.0 est.	Yes	0.014
RTX 3090 Ti 24GB	24	FP16	22.0 est.	Yes	0.011
Mac mini M4 Pro 64GB	64	Q4	21.7 est.	Yes	0.01
RTX 4060 8GB	8	Q4	21.6 est.	Yes	0.072
RX 7900 XTX 24GB	24	FP16	21.0 est.	Yes	0.021
RTX 6000 Ada 48GB	48	FP16	21.0 est.	Yes	0.003
Mac mini M4 32GB	32	Q4	21.0 measured	Yes	0.021
RTX 5060 Ti 16GB	16	FP16	20.7 est.	No	0.048
RTX 3090 24GB	24	FP16	20.5 est.	Yes	0.014
L40S 48GB	48	FP16	18.9 est.	Yes	0.002
Mac Studio M3 Ultra 96GB	96	FP16	17.9 est.	Yes	0.004
Mac Studio M2 Ultra 192GB	192	FP16	17.5 est.	Yes	0.003
RTX 5070 12GB	12	FP16	17.5 est.	No	0.032
RX 7900 XT 20GB	20	FP16	17.5 est.	Yes	0.019
RTX A6000 48GB	48	FP16	16.8 est.	Yes	0.004
T4 16GB	16	FP16	14.8 est.	No	0.007
RTX 4070 Ti 12GB	12	FP16	14.4 est.	No	0.018
RTX 3080 10GB	10	FP16	13.7 est.	No	0.02
RTX 4060 Ti 16GB	16	FP16	13.3 est.	No	0.027
RX 7600 XT 16GB	16	FP16	13.3 est.	No	0.04
RTX 4070 12GB	12	FP16	13.1 est.	No	0.022
RTX 4070 Super 12GB	12	FP16	13.1 est.	No	0.022
MacBook Pro M4 Max 128GB	128	FP16	11.9 est.	Yes	0.003
Arc B580 12GB	12	FP16	11.9 est.	No	0.048
RX 7700 XT 12GB	12	FP16	11.2 est.	No	0.025
RTX 3060 12GB	12	FP16	9.4 est.	No	0.029
MacBook Pro M3 Max 128GB	128	FP16	8.8 est.	Yes	0.002
Arc B570 10GB	10	FP16	6.9 est.	No	0.032
Mac mini M4 Pro 64GB	64	FP16	6.0 est.	Yes	0.003
RTX 5060 8GB	8	FP16	5.2 est.	No	0.017
RTX 5060 Ti 8GB	8	FP16	5.2 est.	No	0.014
RTX 3070 8GB	8	FP16	5.2 est.	No	0.01
RTX 4060 Ti 8GB	8	FP16	3.3 est.	No	0.008
RX 7600 8GB	8	FP16	3.3 est.	No	0.012
RTX 4060 8GB	8	FP16	3.1 est.	No	0.01
Mac mini M4 32GB	32	FP16	2.6 est.	Yes	0.003

Fit checks

RTX 5090 32GB — fits · ≈214.3 tok/s Q4 measured
RTX 3090 Ti 24GB — fits · ≈80.2 tok/s Q4 est.
RTX 4090 24GB — fits · ≈145.0 tok/s Q4 measured
RTX 5080 16GB — fits · ≈136.2 tok/s Q4 measured
RTX 6000 Ada 48GB — fits · ≈138.6 tok/s Q4 measured
RX 7900 XTX 24GB — fits · ≈65.0 tok/s Q4 measured
RTX 3090 24GB — fits · ≈92.0 tok/s Q4 measured
RTX 5070 Ti 16GB — fits · ≈131.9 tok/s Q4 measured
L40S 48GB — fits · ≈68.7 tok/s Q4 est.
Mac Studio M3 Ultra 96GB — fits · ≈65.1 tok/s Q4 est.
Mac Studio M2 Ultra 192GB — fits · ≈63.6 tok/s Q4 est.
RX 7900 XT 20GB — fits · ≈63.6 tok/s Q4 est.

12 popular retail GPUs; every row in the table above also links its Fits? verdict to the full fit check.

Same family: Llama

Llama 3.1 Llama-3.1-70B-Instruct
Llama 3.3 Llama-3.3-70B-Instruct
Llama 3.2 Llama-3.2-3B-Instruct
Llama 4 Llama-4-Scout-17B-16E-Instruct

All Llama models →