gemma-4-31B-it-FP8-block — which GPUs fit

RedHatAI · 31.3B · ctx NoneK · apache-2.0 · HuggingFace

gemma-4-31B-it-FP8-block (RedHatAI) is a 31.3B parameters open-weight model tracked on AI Hashrate. At 8K context we estimate roughly 24.48 GB VRAM for Q4 (weights 17.22 GB + KV 6.26 GB + 1 GB overhead) and 69.86 GB for FP16. About 24 GPUs/accelerators in our catalog fully fit this model at Q4 under that context assumption. Fastest Q4 configs: B200 192GB SXM ≈ 162.6 tok/s (estimated); MI300X 192GB ≈ 107.8 tok/s (estimated); H200 141GB SXM5 ≈ 97.6 tok/s (estimated); H100 80GB SXM5 ≈ 68.1 tok/s (estimated); Gaudi 3 128GB ≈ 66.6 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.

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

Hardware	VRAM	Quant	tok/s	Fits?	tok/s/$
B200 192GB SXM	192	Q4	162.6 est.	Yes	0.004
MI300X 192GB	192	Q4	107.8 est.	Yes	0.007
H200 141GB SXM5	141	Q4	97.6 est.	Yes	0.003
H100 80GB SXM5	80	Q4	68.1 est.	Yes	0.002
Gaudi 3 128GB	128	Q4	66.6 est.	Yes	0.005
RTX 4090 24GB	24	Q4	56.2 est.	No	0.035
RTX 4090D 24GB (China)	24	Q4	56.2 est.	No	0.035
RTX 3090 Ti 24GB	24	Q4	56.2 est.	No	0.028
RX 7900 XTX 24GB	24	Q4	53.5 est.	No	0.054
RTX 3090 24GB	24	Q4	52.2 est.	No	0.035
B200 192GB SXM	192	FP16	44.7 est.	Yes	0.001
A100 80GB SXM4	80	Q4	41.5 est.	Yes	0.003
RTX 5090 32GB	32	Q4	36.4 est.	Yes	0.018
RX 7900 XT 20GB	20	Q4	30.9 est.	No	0.034
MI300X 192GB	192	FP16	29.6 est.	Yes	0.002
H200 141GB SXM5	141	FP16	26.8 est.	Yes	0.001
RTX 5080 16GB	16	Q4	23.7 est.	No	0.024
V100 16GB PCIe	16	Q4	22.3 est.	No	0.003
RTX 5070 Ti 16GB	16	Q4	22.2 est.	No	0.03
RTX 6000 Ada 48GB	48	Q4	19.5 est.	Yes	0.003
H100 80GB SXM5	80	FP16	18.7 est.	Yes	0.001
V100 32GB SXM2	32	Q4	18.3 est.	Yes	0.002
Gaudi 3 128GB	128	FP16	18.3 est.	Yes	0.001
RTX 4080 Super 16GB	16	Q4	18.2 est.	No	0.018
L40S 48GB	48	Q4	17.6 est.	Yes	0.002
Mac Studio M3 Ultra 96GB	96	Q4	16.7 est.	Yes	0.004
RTX 4070 Ti Super 16GB	16	Q4	16.6 est.	No	0.021
Mac Studio M2 Ultra 192GB	192	Q4	16.3 est.	Yes	0.002
RX 9070 16GB	16	Q4	15.8 est.	No	0.029
RX 9070 XT 16GB	16	Q4	15.8 est.	No	0.026
RTX A6000 48GB	48	Q4	15.6 est.	Yes	0.003
RX 7800 XT 16GB	16	Q4	15.4 est.	No	0.031
RX 7900 GRE 16GB	16	Q4	14.2 est.	No	0.026
Arc A770 16GB	16	Q4	13.9 est.	No	0.04
Arc Pro B70 32GB	32	Q4	12.4 est.	Yes	0.012
A100 80GB SXM4	80	FP16	11.4 est.	Yes	0.001
MacBook Pro M4 Max 128GB	128	Q4	11.1 est.	Yes	0.003
RTX 5060 Ti 16GB	16	Q4	11.1 est.	No	0.026
DGX Spark (GH200 480GB)	480	Q4	10.2 est.	Yes	0.003
RTX 5070 12GB	12	Q4	9.3 est.	No	0.017
Titan V 12GB	12	Q4	9.1 est.	No	0.003
MacBook Pro M1 Max 32GB	32	Q4	8.1 est.	Yes	0.002
MacBook Pro M1 Max 64GB	64	Q4	8.1 est.	Yes	0.002
MacBook Pro M3 Max 128GB	128	Q4	8.1 est.	Yes	0.002
T4 16GB	16	Q4	7.9 est.	No	0.004
RTX 4070 Ti 12GB	12	Q4	7.7 est.	No	0.01
Titan Xp 12GB	12	Q4	7.6 est.	No	0.006
RTX 3080 10GB	10	Q4	7.3 est.	No	0.01
RTX 6000 Ada 48GB	48	FP16	7.2 est.	No	0.001
RTX 4060 Ti 16GB	16	Q4	7.1 est.	No	0.014
RX 7600 XT 16GB	16	Q4	7.1 est.	No	0.022
RTX 4070 12GB	12	Q4	7.0 est.	No	0.012
RTX 4070 Super 12GB	12	Q4	7.0 est.	No	0.012
Titan X Pascal 12GB	12	Q4	6.7 est.	No	0.006
L40S 48GB	48	FP16	6.5 est.	No	0.001
Arc B580 12GB	12	Q4	6.3 est.	No	0.025
RTX 5090 32GB	32	FP16	6.0 est.	No	0.003
RX 7700 XT 12GB	12	Q4	6.0 est.	No	0.013
RTX A6000 48GB	48	FP16	5.8 est.	No	0.001
Mac mini M4 Pro 64GB	64	Q4	5.6 est.	Yes	0.003
MacBook Pro M1 Max 64GB	64	FP16	5.4 est.	No	0.001
RTX 3060 12GB	12	Q4	5.0 est.	No	0.015
Mac Studio M3 Ultra 96GB	96	FP16	4.6 est.	Yes	0.001
Mac Studio M2 Ultra 192GB	192	FP16	4.5 est.	Yes	0.001
Ryzen AI Max+ 395 128GB	96	Q4	4.4 est.	Yes	0.002
Ryzen AI Max+ 395 64GB	48	Q4	4.4 est.	Yes	0.003
Ryzen AI Max 390 64GB	48	Q4	4.4 est.	Yes	0.003
Arc B570 10GB	10	Q4	3.7 est.	No	0.017
Mac mini M4 Pro 64GB	64	FP16	3.7 est.	No	0.002
MacBook Pro M4 Max 128GB	128	FP16	3.1 est.	Yes	0.001
V100 32GB SXM2	32	FP16	3.0 est.	No	0.0
RTX 5060 8GB	8	Q4	2.8 est.	No	0.009
RTX 5060 Ti 8GB	8	Q4	2.8 est.	No	0.007
RTX 3070 8GB	8	Q4	2.8 est.	No	0.006
DGX Spark (GH200 480GB)	480	FP16	2.8 est.	Yes	0.001
Mac mini M4 32GB	32	Q4	2.4 est.	Yes	0.002
MacBook Pro M3 Max 128GB	128	FP16	2.2 est.	Yes	0.001
Arc Pro B70 32GB	32	FP16	2.0 est.	No	0.002
RTX 4090 24GB	24	FP16	1.9 est.	No	0.001
RTX 4090D 24GB (China)	24	FP16	1.9 est.	No	0.001
RTX 3090 Ti 24GB	24	FP16	1.9 est.	No	0.001
RTX 3090 24GB	24	FP16	1.8 est.	No	0.001
RX 7900 XTX 24GB	24	FP16	1.8 est.	No	0.002
RTX 4060 Ti 8GB	8	Q4	1.8 est.	No	0.005
RX 7600 8GB	8	Q4	1.8 est.	No	0.007
RTX 4060 8GB	8	Q4	1.7 est.	No	0.006
Ryzen AI Max+ 395 64GB	48	FP16	1.6 est.	No	0.001
Ryzen AI Max 390 64GB	48	FP16	1.6 est.	No	0.001
MacBook Pro M1 Max 32GB	32	FP16	1.3 est.	No	0.0
Tesla P4 8GB	8	Q4	1.2 est.	No	0.015
Ryzen AI Max+ 395 128GB	96	FP16	1.2 est.	Yes	0.001
RX 7900 XT 20GB	20	FP16	1.0 est.	No	0.001
Mac mini M4 32GB	32	FP16	0.4 est.	No	0.0

Fit checks

RTX 5090 32GB — fits · ≈36.4 tok/s Q4 est.
RTX 3090 Ti 24GB — does not fit · ≈56.2 tok/s Q4 est.
RTX 4090 24GB — does not fit · ≈56.2 tok/s Q4 est.
RTX 5080 16GB — does not fit · ≈23.7 tok/s Q4 est.
RTX 6000 Ada 48GB — fits · ≈19.5 tok/s Q4 est.
RX 7900 XTX 24GB — does not fit · ≈53.5 tok/s Q4 est.
RTX 3090 24GB — does not fit · ≈52.2 tok/s Q4 est.
RTX 5070 Ti 16GB — does not fit · ≈22.2 tok/s Q4 est.
L40S 48GB — fits · ≈17.6 tok/s Q4 est.
Mac Studio M3 Ultra 96GB — fits · ≈16.7 tok/s Q4 est.
Mac Studio M2 Ultra 192GB — fits · ≈16.3 tok/s Q4 est.
RX 7900 XT 20GB — does not fit · ≈30.9 tok/s Q4 est.

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