Google Gemma 4

Gemma 4 31B hardware requirements

Fits on one Instinct MI355X OAM at BF16. For 1,000 tokens/s at 32k context you need 6 x H100 SXM5 80GB at $10.98/hr or 12 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $6.00/hr.

Parameters

31.3B

Active per token

All

dense model

KV cache per token

440 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Gemma 4 31B

The strongest part each vendor offers for Gemma 4 31B at BF16 with 32,768 tokens per user, rentable parts first. Each card is one replica, the fewest accelerators that hold the model, and the line it draws on the graph below. Select a card to plot it.

Tokens per second as users grow

Each line is one replica, the fewest accelerators that hold the model, serving more users at once. It climbs while memory bandwidth is the limit, flattens once compute is, and stops where memory runs out at 32,768 tokens per user. Hover a line to bring it forward.

1 x GB300 NVL72 GPU 288GB 1 x Instinct MI355X OAM 1 x Ascend 910C 1 x Gaudi 2 96GB 1 x TPU v7 192GB 4 x Inferentia2

Memory fills at: 1 x GB300 NVL72 GPU 288GB 121 users, 1 x Instinct MI355X OAM 121 users, 1 x Ascend 910C 34 users, 1 x Gaudi 2 96GB 16 users, 1 x TPU v7 192GB 69 users, 4 x Inferentia2 32 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 93.7
Per hour: $5.99
Per million tokens: $17.75
Memory used: 23%

Rent on Daytona

Cheapest per hour

1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 4,096 tokens, one stream, live price

Tokens/s: 20.9
Per hour: $0.50
Per million tokens: $6.64
Memory used: 69%

Rent on RunPod

Cheapest per million tokens

2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 185
Per hour: $1.00
Per million tokens: $1.50
Memory used: 65%

Rent on RunPod

Configurations

Every datacenter GPU in the catalog, sized for these settings. Change precision, context and concurrency and the whole page follows. Each point is one replica: the smallest tensor-parallel group of that GPU that holds the model.

Precision Context Streams

Vendor

Priced only

Priced configurations only; 67 more fit without a live price and appear in the table. Marker size grows with the tensor-parallel width. Roofline estimates.

#									Rent
1	Instinct MI455X	1	29 %	1.2k tok/s	147 tok/s	—	—	—	Specs →
2	Rubin SXM	1	43 %	956 tok/s	139 tok/s	—	—	—	Specs →
3	TPU 8t announced	1	57 %	~555 tok/s	41.2 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	43 %	~533 tok/s	54.2 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	43 %	520 tok/s	50.5 tok/s	$5.99	$3.20	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	43 %	517 tok/s	50.5 tok/s	$12.42	$6.67	—	Rent →
7	GB200 NVL72 GPU 186GB	1	67 %	517 tok/s	50.5 tok/s	$10.50	$5.64	—	Rent →
8	Instinct MI350X OAM	1	43 %	490 tok/s	50.5 tok/s	$6.16	$3.49	1.0 kW	Rent →
9	TPU v7 192GB	1	64 %	477 tok/s	46.5 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	29 %	~610 tok/s	147 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	47 %	481 tok/s	50.5 tok/s	$4.99	$2.88	1.4 kW	Rent →
12	B200 SXM 180GB	1	69 %	476 tok/s	48.6 tok/s	$4.09	$2.39	1.0 kW	Rent →
13	Ascend 970 announced	1	43 %	504 tok/s	90.8 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	1	64 %	401 tok/s	50.5 tok/s	—	—	0.7 kW	Specs →
15	Jalapeño announced	1	57 %	~444 tok/s	97.1 tok/s	—	—	0.7 kW	Specs →
16	Instinct MI325X OAM	1	48 %	300 tok/s	37.8 tok/s	$3.80	$3.52	1.0 kW	Rent →
17	Instinct MI300X 192GB	1	64 %	290 tok/s	33.4 tok/s	$2.39	$2.29	0.8 kW	Rent →
18	BR100	2	98 %	301 tok/s	19.2 tok/s	—	—	1.1 kW	Specs →
19	Ascend 960 announced	1	43 %	263 tok/s	60.5 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	1	88 %	230 tok/s	30.3 tok/s	$2.48	$3.00	0.7 kW	Rent →
21	H100 SXM5 80GB	2	78 %	396 tok/s	40.1 tok/s	$3.66	$2.57	1.4 kW	Rent →
22	TPU v6e 32GB	4	100 %	545 tok/s	37.2 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	1	97 %	184 tok/s	20.7 tok/s	—	—	0.6 kW	Specs →
24	H200 NVL 141GB	1	88 %	~202 tok/s	30.3 tok/s	$3.61	$4.98	0.6 kW	Rent →
25	H100 NVL 94GB	2	66 %	~365 tok/s	46.7 tok/s	$6.38	$4.85	0.8 kW	Rent →
26	Ascend 910C	1	97 %	177 tok/s	20.2 tok/s	—	—	0.6 kW	Specs →
27	H100 PCIe 80GB	2	78 %	279 tok/s	24.0 tok/s	$5.00	$4.98	0.7 kW	Rent →
28	H800 PCIe 80GB	2	78 %	279 tok/s	24.0 tok/s	—	—	0.7 kW	Specs →
29	Ascend 950DT	1	86 %	138 tok/s	25.2 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	2	65 %	205 tok/s	21.5 tok/s	$3.60	$4.89	1.2 kW	Rent →
31	Ascend 950PR	1	97 %	103 tok/s	10.1 tok/s	—	—	0.9 kW	Specs →
32	MTT S5000	2	78 %	~197 tok/s	19.2 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	2	65 %	196 tok/s	19.1 tok/s	$1.38	$1.95	1.2 kW	Rent →
34	Gaudi 3 128GB	1	97 %	303 tok/s	23.3 tok/s	—	—	0.9 kW	Specs →
35	TPU v5p 95GB	2	66 %	212 tok/s	33.1 tok/s	—	—	0.9 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	2	65 %	185 tok/s	21.5 tok/s	$1.00	$1.50	0.6 kW	Rent →
37	Cloud AI 100 Ultra	1	97 %	~60.3 tok/s	3.5 tok/s	—	—	0.1 kW	Specs →
38	Gaudi 2 96GB	2	65 %	197 tok/s	29.4 tok/s	$2.20	$3.10	1.2 kW	Rent →
39	Atlas 350	2	55 %	169 tok/s	16.8 tok/s	—	—	1.2 kW	Specs →
40	Data Center GPU Max 1100 48GB	4	66 %	304 tok/s	27.9 tok/s	—	—	1.2 kW	Specs →
41	GeForce RTX 5090 32GB	4	100 %	195 tok/s	40.7 tok/s	$4.76	$6.77	2.3 kW	Rent →
42	Ascend 910B	2	98 %	168 tok/s	19.2 tok/s	—	—	0.8 kW	Specs →
43	MI250X 128GB	1	97 %	99.2 tok/s	20.7 tok/s	—	—	0.6 kW	Specs →
44	RTX 6000 Ada 48GB	4	66 %	~254 tok/s	21.8 tok/s	$3.88	$4.24	1.2 kW	Rent →
45	Instinct MI250 128GB	1	97 %	94.5 tok/s	20.7 tok/s	—	—	0.5 kW	Specs →
46	L40S 48GB	4	66 %	243 tok/s	19.6 tok/s	$3.04	$3.47	1.4 kW	Rent →
47	A800 PCIe 80GB	2	78 %	145 tok/s	23.2 tok/s	—	—	0.6 kW	Specs →
48	A800 SXM4	2	78 %	147 tok/s	24.4 tok/s	—	—	0.8 kW	Specs →
49	A100 PCIe 80GB	2	78 %	145 tok/s	23.2 tok/s	$2.70	$5.17	0.6 kW	Rent →
50	A100 SXM4 80GB	2	78 %	147 tok/s	24.4 tok/s	$2.80	$5.30	0.8 kW	Rent →
51	A100 PCIe 40GB	4	79 %	263 tok/s	35.3 tok/s	—	—	1.0 kW	Specs →
52	A100 SXM4 40GB	4	79 %	263 tok/s	35.3 tok/s	$6.24	$6.61	1.6 kW	Rent →
53	GeForce RTX 5090 D V2 24GB	8	67 %	~461 tok/s	57.6 tok/s	—	—	4.6 kW	Specs →
54	RTX 5880 Ada 48GB	4	66 %	~212 tok/s	21.8 tok/s	—	—	1.1 kW	Specs →
55	TPU v4 32GB	4	100 %	224 tok/s	27.2 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB	4	100 %	~170 tok/s	13.1 tok/s	$3.32	$5.43	1.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	4	66 %	~219 tok/s	30.5 tok/s	$3.84	$4.86	1.2 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	2	87 %	~116 tok/s	16.1 tok/s	—	—	0.6 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition	4	100 %	161 tok/s	18.2 tok/s	—	—	0.7 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB	4	100 %	166 tok/s	20.3 tok/s	$2.88	$4.82	0.8 kW	Rent →
61	Radeon AI PRO R9700 32GB	4	100 %	144 tok/s	14.5 tok/s	—	—	1.2 kW	Specs →
62	Inferentia2	4	100 %	154 tok/s	18.6 tok/s	—	—	—	Specs →
63	Trainium	4	100 %	154 tok/s	18.6 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB	4	100 %	90.9 tok/s	27.9 tok/s	—	—	1.2 kW	Specs →
65	Arc Pro B70 32GB	4	100 %	~138 tok/s	13.8 tok/s	—	—	0.9 kW	Specs →
66	L40 48GB	4	66 %	151 tok/s	19.6 tok/s	$3.28	$6.04	1.2 kW	Rent →
67	Instinct MI210 PCIe	2	98 %	89.8 tok/s	19.6 tok/s	—	—	0.6 kW	Specs →
68	GeForce RTX 4090 24GB	8	67 %	274 tok/s	43.2 tok/s	$7.12	$7.21	3.6 kW	Rent →
69	A30 24GB	8	67 %	270 tok/s	40.0 tok/s	—	—	1.3 kW	Specs →
70	RTX 4500 Ada 24GB	8	67 %	~211 tok/s	18.5 tok/s	—	—	1.7 kW	Specs →
71	RTX A6000 48GB	4	66 %	130 tok/s	17.4 tok/s	$2.00	$4.27	1.2 kW	Rent →
72	A40 48GB	4	66 %	124 tok/s	15.8 tok/s	$2.36	$5.29	1.2 kW	Rent →
73	H20 141GB HBM3e	1	88 %	42.7 tok/s	25.2 tok/s	—	—	0.4 kW	Specs →
74	H20 96GB	2	65 %	81.1 tok/s	47.9 tok/s	—	—	0.8 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB	8	67 %	~229 tok/s	28.8 tok/s	$4.56	$5.53	1.2 kW	Rent →
76	BI-V100 Tiangai 100	4	100 %	136 tok/s	27.2 tok/s	—	—	1.0 kW	Specs →
77	Atlas 300I Duo 48GB	4	66 %	101 tok/s	9.3 tok/s	—	—	0.6 kW	Specs →
78	Atlas 300I Duo 96GB	2	65 %	53.4 tok/s	4.9 tok/s	—	—	0.3 kW	Specs →
79	RTX A5500 24GB	8	67 %	~223 tok/s	32.9 tok/s	—	—	1.8 kW	Specs →
80	Wormhole n300s	8	67 %	202 tok/s	24.7 tok/s	—	—	2.4 kW	Specs →
81	V100S PCIe 32GB	4	100 %	122 tok/s	25.7 tok/s	$3.52	$8.05	1.0 kW	Rent →
82	ATOM-Max	2	98 %	62.3 tok/s	12.3 tok/s	—	—	0.7 kW	Specs →
83	A10 24GB	8	67 %	197 tok/s	25.7 tok/s	$8.81	$12.43	1.2 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	4	66 %	111 tok/s	19.6 tok/s	—	—	1.2 kW	Specs →
85	Radeon PRO W7900 48GB	4	66 %	111 tok/s	19.6 tok/s	—	—	1.2 kW	Specs →
86	Radeon RX 7900 XTX 24GB	8	67 %	213 tok/s	41.2 tok/s	—	—	2.8 kW	Specs →
87	L4 24GB	8	67 %	156 tok/s	12.9 tok/s	$4.72	$8.42	0.6 kW	Rent →
88	L20 48GB	4	66 %	108 tok/s	19.6 tok/s	—	—	1.1 kW	Specs →
89	Radeon AI PRO R9600D 32GB	4	100 %	88.0 tok/s	14.5 tok/s	—	—	0.6 kW	Specs →
90	Arc Pro B65 32GB	4	100 %	~86.8 tok/s	13.8 tok/s	—	—	0.8 kW	Specs →
91	Arc Pro B60 24GB	8	67 %	~154 tok/s	19.6 tok/s	—	—	1.6 kW	Specs →
92	L2 24GB	8	67 %	135 tok/s	12.9 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB	8	67 %	~149 tok/s	18.5 tok/s	—	—	0.6 kW	Specs →
94	MLU370-X8	4	66 %	85.2 tok/s	13.9 tok/s	—	—	1.0 kW	Specs →
95	Radeon PRO W7800 48GB	4	66 %	85.6 tok/s	19.6 tok/s	—	—	1.0 kW	Specs →
96	Radeon PRO W7800 32GB	4	100 %	80.1 tok/s	13.1 tok/s	—	—	1.0 kW	Specs →
97	GeForce RTX 3090 24GB	8	67 %	132 tok/s	40.1 tok/s	$4.00	$8.42	2.8 kW	Rent →
98	A10G 24GB	8	67 %	123 tok/s	25.7 tok/s	$8.05	$18.14	2.4 kW	Rent →
99	Atlas 300V Pro	4	66 %	50.6 tok/s	4.6 tok/s	—	—	—	Specs →
100	Atlas 300I Pro	8	67 %	95.6 tok/s	8.8 tok/s	—	—	—	Specs →
101	Atlas 300V	8	67 %	75.7 tok/s	8.8 tok/s	—	—	0.6 kW	Specs →
102	GeForce RTX 3090 Ti 24GB	8	67 %	78.2 tok/s	43.2 tok/s	$3.68	$13.08	3.6 kW	Rent →
103	Radeon PRO W6800 32GB	4	100 %	35.4 tok/s	11.6 tok/s	—	—	1.0 kW	Specs →
104	RTX A5000 24GB	8	67 %	54.6 tok/s	32.9 tok/s	$2.16	$10.99	1.8 kW	Rent →
105	Radeon PRO V710 28GB	8	57 %	52.4 tok/s	19.2 tok/s	—	—	1.3 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 185 tokens/s for $1.00/hr; the AMD pick is 1 x Instinct MI300X 192GB at about 290 tokens/s for $2.39/hr. Per rental dollar NVIDIA delivers 1.53x the tokens of AMD here.

NVIDIA 51 parts fit

Best value: 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 185 tok/s for $1.00/hr

Fewest GPUs: 1 x Rubin SXM, 956 tok/s

AMD 18 parts fit

Best value: 1 x Instinct MI300X 192GB, 290 tok/s for $2.39/hr

Fewest GPUs: 1 x Instinct MI455X, 1.2k tok/s

Huawei 10 parts fit

Fewest GPUs: 1 x Ascend 910C, 177 tok/s

Intel 7 parts fit

Best value: 2 x Gaudi 2 96GB, 197 tok/s for $2.20/hr

Fewest GPUs: 1 x Gaudi 3 128GB, 303 tok/s

Google 4 parts fit

Fewest GPUs: 1 x TPU v7 192GB, 477 tok/s

AWS 2 parts fit

Fewest GPUs: 4 x Inferentia2, 154 tok/s

Other 7 parts fit

Fewest GPUs: 1 x Cloud AI 100 Ultra, 60.3 tok/s

Fleet what-if

Compare whole fleets at these settings: a thousand of one part against a hundred of another. Each fleet splits into replicas of its tensor-parallel width; GPUs left over sit idle.

A GPUs Part

B GPUs Part

Fleet letter	Fleet	Replicas	Idle	Tokens/s	Per hour	Power	Per M tokens
A	1,000 x H200 SXM 141GB	1,000 x TP1	0	230k tok/s	$2,480	700 kW	$3.00
B	100 x GB200 NVL72 GPU 186GB	100 x TP1	0	52k tok/s	$1,050	—	$5.64

Size Gemma 4 31B for a tokens per second target

The calculator starts from your traffic instead of a fleet: set a target rate and read the replica count for every GPU.

Open in the LLM GPU Calculator

Frequently asked questions

How much GPU memory does Gemma 4 31B need?

The weights take 62.5 GB at the native BF16 precision (62.5 GB at BF16, 31.3 GB at FP8, 17.6 GB at INT4). Each concurrent stream adds 440 KB of KV cache per token: 0.6 GB at 4,096 tokens and 1.8 GB at 32,768 tokens.

What is the bare minimum to run Gemma 4 31B?

1 x Instinct MI355X OAM at BF16 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $0.50 per hour.

How many H100s do you need to run Gemma 4 31B?

one H100 at BF16 for 4,096 tokens of context and one stream. For 32,768 tokens and 32 concurrent streams the tensor-parallel group is 2 H100s, and 1,000 tokens per second takes 6 H100s across 3 replicas.

How much does Gemma 4 31B cost per million tokens on an H100?

About $2.57 per million output tokens at BF16, 32,768 tokens of context and 32 streams, using the cheapest live on-demand price of $1.83 per GPU-hour. The cheapest part per token is 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.50.

What precision does Gemma 4 31B ship in?

The published checkpoint is BF16. FP8 and INT4 figures describe post-training quantisations that halve and quarter the footprint at a small accuracy cost.

All figures are roofline estimates from the model's config.json and each GPU's published memory, bandwidth and tensor-core peak, with fixed efficiency factors. Prices are the cheapest live on-demand listing per GPU when the site has one. They are a planning floor; a tuned serving stack can do better. Ten global layers use four 512-wide heads; the other fifty use sixteen 256-wide heads over a 1,024-token window. K and V share one tensor.