gpt-oss-20b — which GPUs fit

OpenAI · 21.5B (active 3.6B) · MoE · ctx 128K · Apache 2.0 · HuggingFace

gpt-oss-20b (OpenAI) is a 21.5B parameters MoE (3.6B active per token) open-weight model tracked on AI Hashrate. At 8K context we estimate roughly 13.55 GB VRAM for Q4 (weights 11.83 GB + KV 0.72 GB + 1 GB overhead) and 44.72 GB for FP16. About 35 GPUs/accelerators in our catalog fully fit this model at Q4 under that context assumption. Fastest Q4 configs: B200 192GB SXM ≈ 1414.1 tok/s (estimated); MI300X 192GB ≈ 936.9 tok/s (estimated); H200 141GB SXM5 ≈ 848.5 tok/s (estimated); H100 80GB SXM5 ≈ 592.2 tok/s (estimated); Gaudi 3 128GB ≈ 579.1 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.

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

Hardware	VRAM	Quant	tok/s	Fits?	tok/s/$
B200 192GB SXM	192	Q4	1414.1 est.	Yes	0.035
MI300X 192GB	192	Q4	936.9 est.	Yes	0.062
H200 141GB SXM5	141	Q4	848.5 est.	Yes	0.024
H100 80GB SXM5	80	Q4	592.2 est.	Yes	0.02
Gaudi 3 128GB	128	Q4	579.1 est.	Yes	0.046
B200 192GB SXM	192	FP16	388.9 est.	Yes	0.01
A100 80GB SXM4	80	Q4	360.4 est.	Yes	0.024
RTX 5090 32GB	32	Q4	282.5 measured	Yes	0.141
RTX 5070 12GB	12	Q4	266.2 est.	No	0.485
MI300X 192GB	192	FP16	257.6 est.	Yes	0.017
H200 141GB SXM5	141	FP16	233.3 est.	Yes	0.007
RTX 4090 24GB	24	Q4	225.2 measured	Yes	0.141
RTX 4070 Ti 12GB	12	Q4	219.4 est.	No	0.275
RTX 3080 10GB	10	Q4	209.1 est.	No	0.299
RTX 5080 16GB	16	Q4	204.9 measured	Yes	0.205
RTX 4070 12GB	12	Q4	199.6 est.	No	0.333
RTX 4070 Super 12GB	12	Q4	199.6 est.	No	0.333
RTX 5070 Ti 16GB	16	Q4	189.5 measured	Yes	0.253
RTX 4080 Super 16GB	16	Q4	186.5 measured	Yes	0.187
Arc B580 12GB	12	Q4	180.6 est.	No	0.725
RTX 4090D 24GB (China)	24	Q4	178.2 est.	Yes	0.111
RTX 3090 Ti 24GB	24	Q4	178.2 est.	Yes	0.089
RX 7700 XT 12GB	12	Q4	171.1 est.	No	0.381
RX 7900 XTX 24GB	24	Q4	169.7 est.	Yes	0.17
RTX 6000 Ada 48GB	48	Q4	169.7 est.	Yes	0.025
H100 80GB SXM5	80	FP16	162.8 est.	Yes	0.005
RTX 3090 24GB	24	Q4	161.8 measured	Yes	0.108
Gaudi 3 128GB	128	FP16	159.2 est.	Yes	0.013
L40S 48GB	48	Q4	152.7 est.	Yes	0.019
Mac Studio M3 Ultra 96GB	96	Q4	144.8 est.	Yes	0.036
RTX 3060 12GB	12	Q4	142.6 est.	No	0.433
RX 7900 XT 20GB	20	Q4	141.4 est.	Yes	0.157
RTX A6000 48GB	48	Q4	135.8 est.	Yes	0.03
RTX 5090 32GB	32	FP16	127.4 est.	No	0.064
RTX 4070 Ti Super 16GB	16	Q4	118.8 est.	Yes	0.149
Mac Studio M2 Ultra 192GB	192	Q4	116.1 measured	Yes	0.017
RX 9070 16GB	16	Q4	113.1 est.	Yes	0.206
RX 9070 XT 16GB	16	Q4	113.1 est.	Yes	0.189
RX 7800 XT 16GB	16	Q4	110.3 est.	Yes	0.221
Arc B570 10GB	10	Q4	104.5 est.	No	0.477
RX 7900 GRE 16GB	16	Q4	101.8 est.	Yes	0.185
A100 80GB SXM4	80	FP16	99.1 est.	Yes	0.007
Arc A770 16GB	16	Q4	99.0 est.	Yes	0.284
MacBook Pro M4 Max 128GB	128	Q4	96.5 est.	Yes	0.022
RTX 5060 Ti 16GB	16	Q4	79.2 est.	Yes	0.185
RTX 5060 8GB	8	Q4	78.9 est.	No	0.264
RTX 5060 Ti 8GB	8	Q4	78.9 est.	No	0.208
RTX 3070 8GB	8	Q4	78.9 est.	No	0.158
MacBook Pro M3 Max 128GB	128	Q4	70.7 est.	Yes	0.017
T4 16GB	16	Q4	56.6 est.	Yes	0.028
RTX 4060 Ti 16GB	16	Q4	50.9 est.	Yes	0.102
RX 7600 XT 16GB	16	Q4	50.9 est.	Yes	0.155
RTX 4060 Ti 8GB	8	Q4	50.7 est.	No	0.127
RX 7600 8GB	8	Q4	50.7 est.	No	0.188
Mac mini M4 Pro 64GB	64	Q4	48.3 est.	Yes	0.022
RTX 4060 8GB	8	Q4	47.9 est.	No	0.16
RTX 6000 Ada 48GB	48	FP16	46.7 est.	Yes	0.007
L40S 48GB	48	FP16	42.0 est.	Yes	0.005
RTX 4090 24GB	24	FP16	40.3 est.	No	0.025
RTX 4090D 24GB (China)	24	FP16	40.3 est.	No	0.025
RTX 3090 Ti 24GB	24	FP16	40.3 est.	No	0.02
Mac Studio M3 Ultra 96GB	96	FP16	39.8 est.	Yes	0.01
Mac Studio M2 Ultra 192GB	192	FP16	38.9 est.	Yes	0.006
RX 7900 XTX 24GB	24	FP16	38.4 est.	No	0.038
RTX 3090 24GB	24	FP16	37.4 est.	No	0.025
RTX A6000 48GB	48	FP16	37.3 est.	Yes	0.008
MacBook Pro M4 Max 128GB	128	FP16	26.5 est.	Yes	0.006
RX 7900 XT 20GB	20	FP16	22.2 est.	No	0.025
Mac mini M4 32GB	32	Q4	21.2 est.	Yes	0.021
MacBook Pro M3 Max 128GB	128	FP16	19.4 est.	Yes	0.005
RTX 5080 16GB	16	FP16	17.1 est.	No	0.017
RTX 5070 Ti 16GB	16	FP16	15.9 est.	No	0.021
Mac mini M4 Pro 64GB	64	FP16	13.3 est.	Yes	0.006
RTX 4080 Super 16GB	16	FP16	13.1 est.	No	0.013
RTX 4070 Ti Super 16GB	16	FP16	11.9 est.	No	0.015
RX 9070 16GB	16	FP16	11.4 est.	No	0.021
RX 9070 XT 16GB	16	FP16	11.4 est.	No	0.019
RX 7800 XT 16GB	16	FP16	11.1 est.	No	0.022
RX 7900 GRE 16GB	16	FP16	10.2 est.	No	0.019
Arc A770 16GB	16	FP16	10.0 est.	No	0.029
Mac mini M4 32GB	32	FP16	8.5 est.	No	0.009
RTX 5060 Ti 16GB	16	FP16	8.0 est.	No	0.019
RTX 5070 12GB	12	FP16	6.7 est.	No	0.012
T4 16GB	16	FP16	5.7 est.	No	0.003
RTX 4070 Ti 12GB	12	FP16	5.5 est.	No	0.007
RTX 4060 Ti 16GB	16	FP16	5.1 est.	No	0.01
RX 7600 XT 16GB	16	FP16	5.1 est.	No	0.016
RTX 4070 12GB	12	FP16	5.0 est.	No	0.008
RTX 4070 Super 12GB	12	FP16	5.0 est.	No	0.008
Arc B580 12GB	12	FP16	4.6 est.	No	0.018
RX 7700 XT 12GB	12	FP16	4.3 est.	No	0.01
RTX 3060 12GB	12	FP16	3.6 est.	No	0.011