Meta Llama 4

Llama 4 Maverick 17B-128E hardware requirements

Needs at least 4 GPUs at BF16 (Instinct MI355X OAM). For 1,000 tokens/s at 32k context you need 16 x H100 SXM5 80GB at $29.28/hr or 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $8.00/hr.

Parameters

402B

Active per token

17.0B

mixture of experts

KV cache per token

192 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Llama 4 Maverick 17B-128E

The strongest part each vendor offers for Llama 4 Maverick 17B-128E 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.

8 x GB200 NVL72 GPU 186GB 4 x Instinct MI355X OAM 8 x Ascend 910C 16 x Gaudi 2 96GB 8 x TPU v7 192GB

Memory fills at: 8 x GB200 NVL72 GPU 186GB 222 users, 4 x Instinct MI355X OAM 108 users, 8 x Ascend 910C 63 users, 16 x Gaudi 2 96GB 235 users, 8 x TPU v7 192GB 238 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

4 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 612
Per hour: $23.96
Per million tokens: $10.87
Memory used: 72%

Rent on Daytona

Cheapest per hour

16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

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

Tokens/s: 455
Per hour: $8.00
Per million tokens: $4.88
Memory used: 55%

Rent on RunPod

Cheapest per million tokens

16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 1.1k
Per hour: $8.00
Per million tokens: $2.00
Memory used: 61%

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; 32 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	4	54 %	4.1k tok/s	217 tok/s	—	—	—	Specs →
2	Rubin SXM	4	81 %	3.6k tok/s	205 tok/s	—	—	—	Specs →
3	TPU 8t announced	8	54 %	~2.8k tok/s	115 tok/s	—	—	—	Specs →
4	TPU 8i announced	4	81 %	~1.7k tok/s	80.0 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	4	81 %	1.6k tok/s	74.4 tok/s	$23.96	$4.11	5.6 kW	Rent →
6	GB300 NVL72 GPU 288GB	4	81 %	1.6k tok/s	74.4 tok/s	$49.68	$8.53	—	Rent →
7	GB200 NVL72 GPU 186GB	8	63 %	3.1k tok/s	141 tok/s	$84.00	$7.64	—	Rent →
8	Instinct MI350X OAM	4	81 %	1.6k tok/s	74.4 tok/s	$24.64	$4.34	4.0 kW	Rent →
9	TPU v7 192GB	8	61 %	2.8k tok/s	130 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	4	54 %	~2.8k tok/s	217 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	4	88 %	1.6k tok/s	74.4 tok/s	$19.96	$3.55	5.6 kW	Rent →
12	B200 SXM 180GB	8	65 %	2.9k tok/s	135 tok/s	$32.72	$3.16	8.0 kW	Rent →
13	Ascend 970 announced	4	81 %	2.1k tok/s	134 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	8	61 %	2.7k tok/s	141 tok/s	—	—	5.6 kW	Specs →
15	Jalapeño announced	8	54 %	~3.7k tok/s	271 tok/s	—	—	5.6 kW	Specs →
16	Instinct MI325X OAM	4	91 %	1.1k tok/s	55.8 tok/s	$15.20	$3.96	4.0 kW	Rent →
17	Instinct MI300X 192GB	8	61 %	1.9k tok/s	93.1 tok/s	$19.12	$2.85	6.0 kW	Rent →
18	BR100 2 nodes	16	92 %	1.3k tok/s	49.6 tok/s	—	—	8.8 kW	Specs →
19	Ascend 960 announced	4	81 %	1.2k tok/s	89.3 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	8	83 %	1.6k tok/s	84.4 tok/s	$19.84	$3.50	5.6 kW	Rent →
21	H100 SXM5 80GB 2 nodes	16	73 %	2.2k tok/s	104 tok/s	$29.28	$3.67	11.2 kW	Rent →
22	Data Center GPU Max 1550 128GB	8	91 %	1.2k tok/s	57.6 tok/s	—	—	4.8 kW	Specs →
23	H200 NVL 141GB	8	83 %	~1.5k tok/s	84.4 tok/s	$28.91	$5.50	4.8 kW	Rent →
24	H100 NVL 94GB 2 nodes	16	62 %	~2.3k tok/s	121 tok/s	$51.04	$6.19	6.4 kW	Rent →
25	Ascend 910C	8	91 %	1.1k tok/s	56.2 tok/s	—	—	4.8 kW	Specs →
26	H100 PCIe 80GB 2 nodes	16	73 %	1.4k tok/s	62.0 tok/s	$40.00	$7.78	5.6 kW	Rent →
27	H800 PCIe 80GB 2 nodes	16	73 %	1.4k tok/s	62.0 tok/s	—	—	5.6 kW	Specs →
28	Ascend 950DT	8	81 %	1.1k tok/s	70.3 tok/s	—	—	—	Specs →
29	RTX PRO 6000 Blackwell Workstation Edition 2 nodes	16	61 %	1.2k tok/s	55.6 tok/s	$28.80	$6.86	9.6 kW	Rent →
30	Ascend 950PR	8	91 %	611 tok/s	28.1 tok/s	—	—	7.2 kW	Specs →
31	MTT S5000 2 nodes	16	73 %	~1.1k tok/s	49.6 tok/s	—	—	—	Specs →
32	RTX PRO 6000 Blackwell Server Edition 2 nodes	16	61 %	1.1k tok/s	49.5 tok/s	$11.04	$2.85	9.6 kW	Rent →
33	Gaudi 3 128GB	8	91 %	1.6k tok/s	65.0 tok/s	—	—	7.2 kW	Specs →
34	TPU v5p 95GB 2 nodes	16	62 %	1.5k tok/s	85.7 tok/s	—	—	7.2 kW	Specs →
35	RTX PRO 6000 Blackwell Max-Q Workstation Edition 2 nodes	16	61 %	1.1k tok/s	55.6 tok/s	$8.00	$2.00	4.8 kW	Rent →
36	Cloud AI 100 Ultra	8	91 %	~260 tok/s	9.6 tok/s	—	—	1.2 kW	Specs →
37	Gaudi 2 96GB 2 nodes	16	61 %	1.3k tok/s	76.0 tok/s	$17.60	$3.69	9.6 kW	Rent →
38	Atlas 350 2 nodes	16	52 %	935 tok/s	43.4 tok/s	—	—	9.6 kW	Specs →
39	Ascend 910B 2 nodes	16	92 %	998 tok/s	49.6 tok/s	—	—	6.4 kW	Specs →
40	MI250X 128GB	8	91 %	814 tok/s	57.6 tok/s	—	—	4.5 kW	Specs →
41	Instinct MI250 128GB	8	91 %	789 tok/s	57.6 tok/s	—	—	4.0 kW	Specs →
42	A800 PCIe 80GB 2 nodes	16	73 %	1.0k tok/s	60.0 tok/s	—	—	4.8 kW	Specs →
43	A800 SXM4 2 nodes	16	73 %	1.0k tok/s	63.2 tok/s	—	—	6.4 kW	Specs →
44	A100 PCIe 80GB 2 nodes	16	73 %	1.0k tok/s	60.0 tok/s	$21.60	$5.99	4.8 kW	Rent →
45	A100 SXM4 80GB 2 nodes	16	73 %	1.0k tok/s	63.2 tok/s	$22.40	$6.04	6.4 kW	Rent →
46	RTX PRO 5000 Blackwell 72GB 2 nodes	16	82 %	~754 tok/s	41.7 tok/s	—	—	4.8 kW	Specs →
47	Instinct MI210 PCIe 2 nodes	16	92 %	696 tok/s	50.8 tok/s	—	—	4.8 kW	Specs →
48	H20 141GB HBM3e	8	83 %	451 tok/s	70.3 tok/s	—	—	3.2 kW	Specs →
49	H20 96GB 2 nodes	16	61 %	796 tok/s	124 tok/s	—	—	6.4 kW	Specs →
50	Atlas 300I Duo 96GB 2 nodes	16	61 %	283 tok/s	12.7 tok/s	—	—	2.4 kW	Specs →
51	ATOM-Max 2 nodes	16	92 %	466 tok/s	31.8 tok/s	—	—	5.6 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 1.1k tokens/s for $8.00/hr; the AMD pick is 8 x Instinct MI300X 192GB at about 1.9k tokens/s for $19.12/hr. Per rental dollar NVIDIA delivers 1.42x the tokens of AMD here.

NVIDIA 22 parts fit

Best value: 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 1.1k tok/s for $8.00/hr

Fewest GPUs: 4 x Rubin SXM, 3.6k tok/s

AMD 8 parts fit

Best value: 8 x Instinct MI300X 192GB, 1.9k tok/s for $19.12/hr

Fewest GPUs: 4 x Instinct MI455X, 4.1k tok/s

Huawei 6 parts fit

Fewest GPUs: 8 x Ascend 910C, 1.1k tok/s

Intel 3 parts fit

Best value: 16 x Gaudi 2 96GB, 1.3k tok/s for $17.60/hr

Fewest GPUs: 8 x Gaudi 3 128GB, 1.6k tok/s

Google 2 parts fit

Fewest GPUs: 8 x TPU v7 192GB, 2.8k tok/s

Other 4 parts fit

Fewest GPUs: 8 x Cloud AI 100 Ultra, 260 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	125 x TP8	0	197k tok/s	$2,480	700 kW	$3.50
B	104 x GB200 NVL72 GPU 186GB	13 x TP8	0	40k tok/s	$1,092	—	$7.64

Size Llama 4 Maverick 17B-128E 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 Llama 4 Maverick 17B-128E need?

The weights take 803 GB at the native BF16 precision (803 GB at BF16, 402 GB at FP8, 226 GB at INT4). Each concurrent stream adds 192 KB of KV cache per token: 0.8 GB at 4,096 tokens and 2.8 GB at 32,768 tokens.

What is the bare minimum to run Llama 4 Maverick 17B-128E?

4 x Instinct MI355X OAM at BF16 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $8.00 per hour.

How many H100s do you need to run Llama 4 Maverick 17B-128E?

16 H100s at BF16 for 4,096 tokens of context and one stream. For 32,768 tokens and 32 concurrent streams the tensor-parallel group is 16 H100s, and 1,000 tokens per second takes 16 H100s across 1 replicas.

How much does Llama 4 Maverick 17B-128E cost per million tokens on an H100?

About $3.67 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 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $2.00.

What precision does Llama 4 Maverick 17B-128E 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. 36 of 48 layers attend within 8,192-token chunks, so only 12 keep the full context; experts sit on every second layer. The Meta repo is gated; geometry read from the unsloth mirror.