Google Gemma 3

Gemma 3 27B hardware requirements

Fits on one Instinct MI355X OAM at BF16. For 1,000 tokens/s at 32k context you need 8 x H100 SXM5 80GB at $14.64/hr or 12 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $6.00/hr.

Parameters

27.4B

Active per token

All

dense model

KV cache per token

496 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Gemma 3 27B

The strongest part each vendor offers for Gemma 3 27B 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 3 128GB 1 x TPU v7 192GB 2 x Inferentia2

Memory fills at: 1 x GB300 NVL72 GPU 288GB 71 users, 1 x Instinct MI355X OAM 71 users, 1 x Ascend 910C 21 users, 1 x Gaudi 3 128GB 21 users, 1 x TPU v7 192GB 41 users, 2 x Inferentia2 1 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 106
Per hour: $5.99
Per million tokens: $15.64
Memory used: 20%

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: 23.7
Per hour: $0.50
Per million tokens: $5.85
Memory used: 61%

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: 187
Per hour: $1.00
Per million tokens: $1.49
Memory used: 84%

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; 68 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	37 %	1.2k tok/s	113 tok/s	—	—	—	Specs →
2	Rubin SXM	1	56 %	987 tok/s	107 tok/s	—	—	—	Specs →
3	TPU 8t announced	1	75 %	~526 tok/s	31.6 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	56 %	~529 tok/s	41.7 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	56 %	513 tok/s	38.8 tok/s	$5.99	$3.25	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	56 %	511 tok/s	38.8 tok/s	$12.42	$6.76	—	Rent →
7	GB200 NVL72 GPU 186GB	1	87 %	511 tok/s	38.8 tok/s	$10.50	$5.71	—	Rent →
8	Instinct MI350X OAM	1	56 %	487 tok/s	38.8 tok/s	$6.16	$3.51	1.0 kW	Rent →
9	TPU v7 192GB	1	84 %	471 tok/s	35.8 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	37 %	~654 tok/s	113 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	61 %	479 tok/s	38.8 tok/s	$4.99	$2.89	1.4 kW	Rent →
12	B200 SXM 180GB	1	90 %	472 tok/s	37.3 tok/s	$4.09	$2.41	1.0 kW	Rent →
13	Ascend 970 announced	1	56 %	530 tok/s	69.8 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	1	84 %	408 tok/s	38.8 tok/s	—	—	0.7 kW	Specs →
15	Jalapeño announced	1	75 %	~473 tok/s	74.6 tok/s	—	—	0.7 kW	Specs →
16	Instinct MI325X OAM	1	63 %	305 tok/s	29.1 tok/s	$3.80	$3.46	1.0 kW	Rent →
17	Instinct MI300X 192GB	1	84 %	292 tok/s	25.7 tok/s	$2.39	$2.27	0.8 kW	Rent →
18	BR100	4	64 %	526 tok/s	27.9 tok/s	—	—	2.2 kW	Specs →
19	Ascend 960 announced	1	56 %	282 tok/s	46.5 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	2	57 %	447 tok/s	44.2 tok/s	$4.96	$3.08	1.4 kW	Rent →
21	H100 SXM5 80GB	4	51 %	745 tok/s	58.5 tok/s	$7.32	$2.73	2.8 kW	Rent →
22	TPU v6e 32GB	8	65 %	961 tok/s	54.0 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	2	63 %	352 tok/s	30.2 tok/s	—	—	1.2 kW	Specs →
24	H200 NVL 141GB	2	57 %	~397 tok/s	44.2 tok/s	$7.23	$5.06	1.2 kW	Rent →
25	H100 NVL 94GB	2	86 %	~373 tok/s	35.9 tok/s	$6.38	$4.76	0.8 kW	Rent →
26	Ascend 910C	2	63 %	338 tok/s	29.5 tok/s	—	—	1.2 kW	Specs →
27	H100 PCIe 80GB	4	51 %	512 tok/s	34.9 tok/s	$10.00	$5.43	1.4 kW	Rent →
28	H800 PCIe 80GB	4	51 %	512 tok/s	34.9 tok/s	—	—	1.4 kW	Specs →
29	Ascend 950DT	2	56 %	276 tok/s	36.8 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	2	84 %	204 tok/s	16.5 tok/s	$3.60	$4.90	1.2 kW	Rent →
31	Ascend 950PR	2	63 %	194 tok/s	14.7 tok/s	—	—	1.8 kW	Specs →
32	MTT S5000	4	51 %	~368 tok/s	27.9 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	2	84 %	194 tok/s	14.7 tok/s	$1.38	$1.98	1.2 kW	Rent →
34	Gaudi 3 128GB	2	63 %	549 tok/s	34.1 tok/s	—	—	1.8 kW	Specs →
35	TPU v5p 95GB	2	85 %	221 tok/s	25.5 tok/s	—	—	0.9 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	2	84 %	187 tok/s	16.5 tok/s	$1.00	$1.49	0.6 kW	Rent →
37	Cloud AI 100 Ultra	2	63 %	~103 tok/s	5.0 tok/s	—	—	0.3 kW	Specs →
38	Gaudi 2 96GB	2	84 %	204 tok/s	22.6 tok/s	—	—	1.2 kW	Specs →
39	Atlas 350	2	72 %	167 tok/s	12.9 tok/s	—	—	1.2 kW	Specs →
40	Data Center GPU Max 1100 48GB	4	85 %	297 tok/s	21.4 tok/s	—	—	1.2 kW	Specs →
41	GeForce RTX 5090 32GB	8	65 %	392 tok/s	59.1 tok/s	$9.52	$6.75	4.6 kW	Rent →
42	Ascend 910B	4	64 %	321 tok/s	27.9 tok/s	—	—	1.6 kW	Specs →
43	MI250X 128GB	2	63 %	200 tok/s	30.2 tok/s	—	—	1.1 kW	Specs →
44	RTX 6000 Ada 48GB	4	85 %	~246 tok/s	16.8 tok/s	$3.96	$4.47	1.2 kW	Rent →
45	Instinct MI250 128GB	2	63 %	191 tok/s	30.2 tok/s	—	—	1.0 kW	Specs →
46	L40S 48GB	4	85 %	234 tok/s	15.1 tok/s	$3.04	$3.62	1.4 kW	Rent →
47	A800 PCIe 80GB	4	51 %	287 tok/s	33.8 tok/s	—	—	1.2 kW	Specs →
48	A800 SXM4	4	51 %	290 tok/s	35.6 tok/s	—	—	1.6 kW	Specs →
49	A100 PCIe 80GB	4	51 %	287 tok/s	33.8 tok/s	$5.40	$5.23	1.2 kW	Rent →
50	A100 SXM4 80GB	4	51 %	290 tok/s	35.6 tok/s	$5.40	$5.16	1.6 kW	Rent →
51	A100 PCIe 40GB	8	51 %	508 tok/s	51.3 tok/s	—	—	2.0 kW	Specs →
52	A100 SXM4 40GB	8	51 %	508 tok/s	51.3 tok/s	$12.49	$6.82	3.2 kW	Rent →
53	GeForce RTX 5090 D V2 24GB	8	87 %	~469 tok/s	44.3 tok/s	—	—	4.6 kW	Specs →
54	RTX 5880 Ada 48GB	4	85 %	~210 tok/s	16.8 tok/s	—	—	1.1 kW	Specs →
55	TPU v4 32GB	8	65 %	429 tok/s	39.6 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB	8	65 %	~306 tok/s	19.0 tok/s	$6.64	$6.03	2.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	4	85 %	~226 tok/s	23.5 tok/s	$3.84	$4.73	1.2 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	4	56 %	~226 tok/s	23.5 tok/s	—	—	1.2 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition	8	65 %	306 tok/s	26.4 tok/s	—	—	1.3 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB	8	65 %	318 tok/s	29.5 tok/s	$5.76	$5.03	1.6 kW	Rent →
61	Radeon AI PRO R9700 32GB	8	65 %	270 tok/s	21.1 tok/s	—	—	2.4 kW	Specs →
62	Inferentia2	8	65 %	295 tok/s	27.0 tok/s	—	—	—	Specs →
63	Trainium	8	65 %	295 tok/s	27.0 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB	8	65 %	187 tok/s	40.5 tok/s	—	—	2.4 kW	Specs →
65	Arc Pro B70 32GB	8	65 %	~259 tok/s	20.0 tok/s	—	—	1.8 kW	Specs →
66	L40 48GB	4	85 %	154 tok/s	15.1 tok/s	$3.28	$5.91	1.2 kW	Rent →
67	Instinct MI210 PCIe	4	64 %	181 tok/s	28.6 tok/s	—	—	1.2 kW	Specs →
68	GeForce RTX 4090 24GB	8	87 %	285 tok/s	33.2 tok/s	$3.20	$3.12	3.6 kW	Rent →
69	A30 24GB	8	87 %	279 tok/s	30.8 tok/s	—	—	1.3 kW	Specs →
70	RTX 4500 Ada 24GB	8	87 %	~205 tok/s	14.2 tok/s	—	—	1.7 kW	Specs →
71	RTX A6000 48GB	4	85 %	133 tok/s	13.4 tok/s	$2.00	$4.17	1.2 kW	Rent →
72	A40 48GB	4	85 %	126 tok/s	12.1 tok/s	$2.36	$5.19	1.2 kW	Rent →
73	H20 141GB HBM3e	2	57 %	90.2 tok/s	36.8 tok/s	—	—	0.8 kW	Specs →
74	H20 96GB	2	84 %	90.2 tok/s	36.8 tok/s	—	—	0.8 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB	8	87 %	~233 tok/s	22.2 tok/s	$4.56	$5.43	1.2 kW	Rent →
76	BI-V100 Tiangai 100	8	65 %	272 tok/s	39.6 tok/s	—	—	2.0 kW	Specs →
77	Atlas 300I Duo 48GB	4	85 %	99.0 tok/s	7.1 tok/s	—	—	0.6 kW	Specs →
78	Atlas 300I Duo 96GB	2	84 %	52.2 tok/s	3.8 tok/s	—	—	0.3 kW	Specs →
79	RTX A5500 24GB	8	87 %	~230 tok/s	25.3 tok/s	—	—	1.8 kW	Specs →
80	Wormhole n300s	8	87 %	205 tok/s	19.0 tok/s	—	—	2.4 kW	Specs →
81	V100S PCIe 32GB	8	65 %	244 tok/s	37.4 tok/s	$7.04	$8.01	2.0 kW	Rent →
82	ATOM-Max	4	64 %	125 tok/s	17.9 tok/s	—	—	1.4 kW	Specs →
83	A10 24GB	8	87 %	201 tok/s	19.8 tok/s	$8.81	$12.16	1.2 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	4	85 %	117 tok/s	15.1 tok/s	—	—	1.2 kW	Specs →
85	Radeon PRO W7900 48GB	4	85 %	117 tok/s	15.1 tok/s	—	—	1.2 kW	Specs →
86	Radeon RX 7900 XTX 24GB	8	87 %	225 tok/s	31.6 tok/s	—	—	2.8 kW	Specs →
87	L4 24GB	8	87 %	150 tok/s	9.9 tok/s	$4.72	$8.73	0.6 kW	Rent →
88	L20 48GB	4	85 %	114 tok/s	15.1 tok/s	—	—	1.1 kW	Specs →
89	Radeon AI PRO R9600D 32GB	8	65 %	174 tok/s	21.1 tok/s	—	—	1.2 kW	Specs →
90	Arc Pro B65 32GB	8	65 %	~171 tok/s	20.0 tok/s	—	—	1.6 kW	Specs →
91	Arc Pro B60 24GB	8	87 %	~157 tok/s	15.0 tok/s	—	—	1.6 kW	Specs →
92	L2 24GB	8	87 %	132 tok/s	9.9 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB	8	87 %	~152 tok/s	14.2 tok/s	—	—	0.6 kW	Specs →
94	MLU370-X8	4	85 %	88.9 tok/s	10.7 tok/s	—	—	1.0 kW	Specs →
95	Radeon PRO W7800 48GB	4	85 %	91.5 tok/s	15.1 tok/s	—	—	1.0 kW	Specs →
96	Radeon PRO W7800 32GB	8	65 %	158 tok/s	19.0 tok/s	—	—	2.1 kW	Specs →
97	GeForce RTX 3090 24GB	8	87 %	143 tok/s	30.9 tok/s	$4.00	$7.75	2.8 kW	Rent →
98	A10G 24GB	8	87 %	131 tok/s	19.8 tok/s	$8.05	$17.06	2.4 kW	Rent →
99	Atlas 300V Pro	4	85 %	49.6 tok/s	3.6 tok/s	—	—	—	Specs →
100	Atlas 300I Pro	8	87 %	93.6 tok/s	6.8 tok/s	—	—	—	Specs →
101	Atlas 300V	8	87 %	76.3 tok/s	6.8 tok/s	—	—	0.6 kW	Specs →
102	GeForce RTX 3090 Ti 24GB	8	87 %	86.7 tok/s	33.2 tok/s	$3.68	$11.79	3.6 kW	Rent →
103	Radeon PRO W6800 32GB	8	65 %	72.9 tok/s	16.9 tok/s	—	—	2.0 kW	Specs →
104	RTX A5000 24GB	8	87 %	60.7 tok/s	25.3 tok/s	$2.16	$9.88	1.8 kW	Rent →
105	Radeon PRO V710 28GB	8	74 %	57.4 tok/s	14.8 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 187 tokens/s for $1.00/hr; the AMD pick is 1 x Instinct MI300X 192GB at about 292 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, 187 tok/s for $1.00/hr

Fewest GPUs: 1 x Rubin SXM, 987 tok/s

AMD 18 parts fit

Best value: 1 x Instinct MI300X 192GB, 292 tok/s for $2.39/hr

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

Huawei 10 parts fit

Fewest GPUs: 2 x Ascend 910C, 338 tok/s

Intel 7 parts fit

Fewest GPUs: 2 x Gaudi 3 128GB, 549 tok/s

Google 4 parts fit

Fewest GPUs: 1 x TPU v7 192GB, 471 tok/s

AWS 2 parts fit

Fewest GPUs: 8 x Inferentia2, 295 tok/s

Other 7 parts fit

Fewest GPUs: 2 x Cloud AI 100 Ultra, 103 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	500 x TP2	0	223k tok/s	$2,480	700 kW	$3.08
B	100 x GB200 NVL72 GPU 186GB	100 x TP1	0	51k tok/s	$1,050	—	$5.71

Size Gemma 3 27B 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 3 27B need?

The weights take 54.9 GB at the native BF16 precision (54.9 GB at BF16, 27.4 GB at FP8, 15.4 GB at INT4). Each concurrent stream adds 496 KB of KV cache per token: 0.8 GB at 4,096 tokens and 3.1 GB at 32,768 tokens.

What is the bare minimum to run Gemma 3 27B?

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 3 27B?

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 4 H100s, and 1,000 tokens per second takes 8 H100s across 2 replicas.

How much does Gemma 3 27B cost per million tokens on an H100?

About $2.73 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 $1.49.

What precision does Gemma 3 27B 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. One layer in six attends globally; the rest use a 1,024-token sliding window. The Google repo is gated; geometry read from the unsloth mirror.