Meta Llama 4

Llama 4 Scout 17B-16E 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

109B

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 Scout 17B-16E

The strongest part each vendor offers for Llama 4 Scout 17B-16E 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.

2 x GB200 NVL72 GPU 186GB 1 x Instinct MI355X OAM 2 x Ascend 950DT 4 x Gaudi 2 96GB 2 x TPU v7 192GB 8 x Inferentia2

Memory fills at: 2 x GB200 NVL72 GPU 186GB 49 users, 1 x Instinct MI355X OAM 21 users, 2 x Ascend 950DT 20 users, 4 x Gaudi 2 96GB 53 users, 2 x TPU v7 192GB 53 users, 8 x Inferentia2 7 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 170
Per hour: $5.99
Per million tokens: $9.78
Memory used: 79%

Rent on Daytona

Cheapest per hour

4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

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

Tokens/s: 137
Per hour: $2.00
Per million tokens: $4.07
Memory used: 59%

Rent on RunPod

Cheapest per million tokens

4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 338
Per hour: $2.00
Per million tokens: $1.64
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; 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	74 %	1.2k tok/s	61.8 tok/s	—	—	—	Specs →
2	Rubin SXM	2	55 %	1.9k tok/s	111 tok/s	—	—	—	Specs →
3	TPU 8t announced	2	74 %	~801 tok/s	32.9 tok/s	—	—	—	Specs →
4	TPU 8i announced	2	55 %	~915 tok/s	43.4 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	2	55 %	871 tok/s	40.3 tok/s	$11.98	$3.82	2.8 kW	Rent →
6	GB300 NVL72 GPU 288GB	2	55 %	869 tok/s	40.3 tok/s	$24.84	$7.94	—	Rent →
7	GB200 NVL72 GPU 186GB	2	86 %	869 tok/s	40.3 tok/s	$21.00	$6.71	—	Rent →
8	Instinct MI350X OAM	2	55 %	846 tok/s	40.3 tok/s	$12.32	$4.04	2.0 kW	Rent →
9	TPU v7 192GB	2	83 %	802 tok/s	37.2 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	74 %	~780 tok/s	61.8 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	2	61 %	839 tok/s	40.3 tok/s	$9.98	$3.31	2.8 kW	Rent →
12	B200 SXM 180GB	2	89 %	818 tok/s	38.8 tok/s	$8.18	$2.78	2.0 kW	Rent →
13	Ascend 970 announced	2	55 %	1.1k tok/s	72.6 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	2	83 %	762 tok/s	40.3 tok/s	—	—	1.4 kW	Specs →
15	Jalapeño announced	2	74 %	~1.0k tok/s	77.6 tok/s	—	—	1.4 kW	Specs →
16	Instinct MI325X OAM	2	62 %	571 tok/s	30.2 tok/s	$7.60	$3.70	2.0 kW	Rent →
17	Instinct MI300X 192GB	2	83 %	530 tok/s	26.7 tok/s	$4.78	$2.51	1.5 kW	Rent →
18	BR100	8	63 %	747 tok/s	28.9 tok/s	—	—	4.4 kW	Specs →
19	Ascend 960 announced	2	55 %	631 tok/s	48.4 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	4	57 %	846 tok/s	45.8 tok/s	$9.92	$3.26	2.8 kW	Rent →
21	H100 SXM5 80GB	8	50 %	1.3k tok/s	60.4 tok/s	$14.64	$3.18	5.6 kW	Rent →
22	TPU v6e 32GB 2 nodes	16	64 %	1.3k tok/s	52.1 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	4	63 %	629 tok/s	31.3 tok/s	—	—	2.4 kW	Specs →
24	H200 NVL 141GB	4	57 %	~784 tok/s	45.8 tok/s	$14.46	$5.12	2.4 kW	Rent →
25	H100 NVL 94GB	4	86 %	~698 tok/s	37.2 tok/s	$12.76	$5.08	1.6 kW	Rent →
26	Ascend 910C	4	63 %	609 tok/s	30.6 tok/s	—	—	2.4 kW	Specs →
27	H100 PCIe 80GB	8	50 %	824 tok/s	36.1 tok/s	$20.00	$6.74	2.8 kW	Rent →
28	H800 PCIe 80GB	8	50 %	824 tok/s	36.1 tok/s	—	—	2.8 kW	Specs →
29	Ascend 950DT	4	56 %	580 tok/s	38.2 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	4	84 %	356 tok/s	17.1 tok/s	$7.20	$5.62	2.4 kW	Rent →
31	Ascend 950PR	4	63 %	329 tok/s	15.3 tok/s	—	—	3.6 kW	Specs →
32	MTT S5000	8	50 %	~622 tok/s	28.9 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	4	84 %	329 tok/s	15.3 tok/s	$2.76	$2.33	2.4 kW	Rent →
34	Gaudi 3 128GB	4	63 %	848 tok/s	35.3 tok/s	—	—	3.6 kW	Specs →
35	TPU v5p 95GB	4	85 %	442 tok/s	26.4 tok/s	—	—	1.8 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	4	84 %	338 tok/s	17.1 tok/s	$2.00	$1.64	1.2 kW	Rent →
37	Cloud AI 100 Ultra	4	63 %	~141 tok/s	5.2 tok/s	—	—	0.6 kW	Specs →
38	Gaudi 2 96GB	4	84 %	403 tok/s	23.4 tok/s	$4.40	$3.04	2.4 kW	Rent →
39	Atlas 350	4	72 %	285 tok/s	13.4 tok/s	—	—	2.4 kW	Specs →
40	Data Center GPU Max 1100 48GB	8	85 %	491 tok/s	22.2 tok/s	—	—	2.4 kW	Specs →
41	GeForce RTX 5090 32GB 2 nodes	16	64 %	795 tok/s	57.1 tok/s	$19.04	$6.65	9.2 kW	Rent →
42	Ascend 910B	8	63 %	575 tok/s	28.9 tok/s	—	—	3.2 kW	Specs →
43	MI250X 128GB	4	63 %	436 tok/s	31.3 tok/s	—	—	2.2 kW	Specs →
44	RTX 6000 Ada 48GB	8	85 %	~396 tok/s	17.3 tok/s	$7.76	$5.44	2.4 kW	Rent →
45	Instinct MI250 128GB	4	63 %	422 tok/s	31.3 tok/s	—	—	2.0 kW	Specs →
46	L40S 48GB	8	85 %	366 tok/s	15.6 tok/s	$6.08	$4.61	2.8 kW	Rent →
47	A800 PCIe 80GB	8	50 %	576 tok/s	34.9 tok/s	—	—	2.4 kW	Specs →
48	A800 SXM4	8	50 %	591 tok/s	36.8 tok/s	—	—	3.2 kW	Specs →
49	A100 PCIe 80GB	8	50 %	576 tok/s	34.9 tok/s	$10.80	$5.21	2.4 kW	Rent →
50	A100 SXM4 80GB	8	50 %	591 tok/s	36.8 tok/s	$11.20	$5.26	3.2 kW	Rent →
51	A100 PCIe 40GB 2 nodes	16	51 %	903 tok/s	49.5 tok/s	—	—	4.0 kW	Specs →
52	A100 SXM4 40GB 2 nodes	16	51 %	903 tok/s	49.5 tok/s	$24.98	$7.69	6.4 kW	Rent →
53	GeForce RTX 5090 D V2 24GB 2 nodes	16	87 %	~812 tok/s	42.8 tok/s	—	—	9.2 kW	Specs →
54	RTX 5880 Ada 48GB	8	85 %	~363 tok/s	17.3 tok/s	—	—	2.3 kW	Specs →
55	TPU v4 32GB 2 nodes	16	64 %	736 tok/s	38.2 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB 2 nodes	16	64 %	~440 tok/s	18.3 tok/s	$13.28	$8.39	4.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	8	85 %	~434 tok/s	24.2 tok/s	$7.68	$4.92	2.4 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	8	56 %	~434 tok/s	24.2 tok/s	—	—	2.4 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition 2 nodes	16	64 %	510 tok/s	25.5 tok/s	—	—	2.6 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB 2 nodes	16	64 %	547 tok/s	28.5 tok/s	$11.52	$5.85	3.2 kW	Rent →
61	Radeon AI PRO R9700 32GB 2 nodes	16	64 %	432 tok/s	20.4 tok/s	—	—	4.8 kW	Specs →
62	Inferentia2 2 nodes	16	64 %	505 tok/s	26.1 tok/s	—	—	—	Specs →
63	Trainium 2 nodes	16	64 %	505 tok/s	26.1 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB 2 nodes	16	64 %	415 tok/s	39.1 tok/s	—	—	4.8 kW	Specs →
65	Arc Pro B70 32GB 2 nodes	16	64 %	~412 tok/s	19.4 tok/s	—	—	3.7 kW	Specs →
66	L40 48GB	8	85 %	289 tok/s	15.6 tok/s	$6.56	$6.30	2.4 kW	Rent →
67	Instinct MI210 PCIe	8	63 %	399 tok/s	29.6 tok/s	—	—	2.4 kW	Specs →
68	GeForce RTX 4090 24GB 2 nodes	16	87 %	533 tok/s	32.1 tok/s	$14.24	$7.41	7.2 kW	Rent →
69	A30 24GB 2 nodes	16	87 %	512 tok/s	29.7 tok/s	—	—	2.6 kW	Specs →
70	RTX 4500 Ada 24GB 2 nodes	16	87 %	~311 tok/s	13.8 tok/s	—	—	3.4 kW	Specs →
71	RTX A6000 48GB	8	85 %	253 tok/s	13.9 tok/s	$4.00	$4.39	2.4 kW	Rent →
72	A40 48GB	8	85 %	236 tok/s	12.6 tok/s	$4.72	$5.56	2.4 kW	Rent →
73	H20 141GB HBM3e	4	57 %	240 tok/s	38.2 tok/s	—	—	1.6 kW	Specs →
74	H20 96GB	4	84 %	240 tok/s	38.2 tok/s	—	—	1.6 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB 2 nodes	16	87 %	~405 tok/s	21.4 tok/s	$9.12	$6.26	2.3 kW	Rent →
76	BI-V100 Tiangai 100 2 nodes	16	64 %	546 tok/s	38.2 tok/s	—	—	4.0 kW	Specs →
77	Atlas 300I Duo 48GB	8	85 %	163 tok/s	7.4 tok/s	—	—	1.2 kW	Specs →
78	Atlas 300I Duo 96GB	4	84 %	86.5 tok/s	3.9 tok/s	—	—	0.6 kW	Specs →
79	RTX A5500 24GB 2 nodes	16	87 %	~422 tok/s	24.5 tok/s	—	—	3.7 kW	Specs →
80	Wormhole n300s 2 nodes	16	87 %	352 tok/s	18.3 tok/s	—	—	4.8 kW	Specs →
81	V100S PCIe 32GB 2 nodes	16	64 %	498 tok/s	36.1 tok/s	$14.08	$7.86	4.0 kW	Rent →
82	ATOM-Max	8	63 %	267 tok/s	18.5 tok/s	—	—	2.8 kW	Specs →
83	A10 24GB 2 nodes	16	87 %	354 tok/s	19.1 tok/s	$17.63	$13.84	2.4 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	8	85 %	242 tok/s	15.6 tok/s	—	—	2.4 kW	Specs →
85	Radeon PRO W7900 48GB	8	85 %	242 tok/s	15.6 tok/s	—	—	2.4 kW	Specs →
86	Radeon RX 7900 XTX 24GB 2 nodes	16	87 %	448 tok/s	30.6 tok/s	—	—	5.7 kW	Specs →
87	L4 24GB 2 nodes	16	87 %	222 tok/s	9.6 tok/s	$9.44	$11.80	1.2 kW	Rent →
88	L20 48GB	8	85 %	238 tok/s	15.6 tok/s	—	—	2.2 kW	Specs →
89	Radeon AI PRO R9600D 32GB 2 nodes	16	64 %	329 tok/s	20.4 tok/s	—	—	2.4 kW	Specs →
90	Arc Pro B65 32GB 2 nodes	16	64 %	~320 tok/s	19.4 tok/s	—	—	3.2 kW	Specs →
91	Arc Pro B60 24GB 2 nodes	16	87 %	~273 tok/s	14.5 tok/s	—	—	3.2 kW	Specs →
92	L2 24GB 2 nodes	16	87 %	208 tok/s	9.6 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB 2 nodes	16	87 %	~262 tok/s	13.8 tok/s	—	—	1.1 kW	Specs →
94	MLU370-X8	8	85 %	180 tok/s	11.1 tok/s	—	—	2.0 kW	Specs →
95	Radeon PRO W7800 48GB	8	85 %	204 tok/s	15.6 tok/s	—	—	2.1 kW	Specs →
96	Radeon PRO W7800 32GB 2 nodes	16	64 %	299 tok/s	18.3 tok/s	—	—	4.2 kW	Specs →
97	GeForce RTX 3090 24GB 2 nodes	16	87 %	318 tok/s	29.8 tok/s	$8.00	$6.99	5.6 kW	Rent →
98	A10G 24GB 2 nodes	16	87 %	266 tok/s	19.1 tok/s	$16.10	$16.82	4.8 kW	Rent →
99	Atlas 300V Pro	8	85 %	81.9 tok/s	3.7 tok/s	—	—	—	Specs →
100	Atlas 300I Pro 2 nodes	16	87 %	145 tok/s	6.5 tok/s	—	—	—	Specs →
101	Atlas 300V 2 nodes	16	87 %	129 tok/s	6.5 tok/s	—	—	1.2 kW	Specs →
102	GeForce RTX 3090 Ti 24GB 2 nodes	16	87 %	213 tok/s	32.1 tok/s	$7.36	$9.59	7.2 kW	Rent →
103	Radeon PRO W6800 32GB 2 nodes	16	64 %	164 tok/s	16.3 tok/s	—	—	4.0 kW	Specs →
104	RTX A5000 24GB 2 nodes	16	87 %	151 tok/s	24.5 tok/s	$4.32	$7.95	3.7 kW	Rent →
105	Radeon PRO V710 28GB 2 nodes	16	74 %	132 tok/s	14.3 tok/s	—	—	2.5 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 338 tokens/s for $2.00/hr; the AMD pick is 2 x Instinct MI300X 192GB at about 530 tokens/s for $4.78/hr. Per rental dollar NVIDIA delivers 1.53x the tokens of AMD here.

NVIDIA 51 parts fit

Best value: 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 338 tok/s for $2.00/hr

Fewest GPUs: 2 x Rubin SXM, 1.9k tok/s

AMD 18 parts fit

Best value: 2 x Instinct MI300X 192GB, 530 tok/s for $4.78/hr

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

Huawei 10 parts fit

Fewest GPUs: 4 x Ascend 910C, 609 tok/s

Intel 7 parts fit

Best value: 4 x Gaudi 2 96GB, 403 tok/s for $4.40/hr

Fewest GPUs: 4 x Gaudi 3 128GB, 848 tok/s

Google 4 parts fit

Fewest GPUs: 2 x TPU v7 192GB, 802 tok/s

AWS 2 parts fit

Fewest GPUs: 16 x Inferentia2, 505 tok/s

Other 7 parts fit

Fewest GPUs: 4 x Cloud AI 100 Ultra, 141 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	250 x TP4	0	211k tok/s	$2,480	700 kW	$3.26
B	100 x GB200 NVL72 GPU 186GB	50 x TP2	0	43k tok/s	$1,050	—	$6.71

Size Llama 4 Scout 17B-16E 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 Scout 17B-16E need?

The weights take 217 GB at the native BF16 precision (217 GB at BF16, 109 GB at FP8, 61.1 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 Scout 17B-16E?

1 x Instinct MI355X OAM at BF16 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $2.00 per hour.

How many H100s do you need to run Llama 4 Scout 17B-16E?

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

How much does Llama 4 Scout 17B-16E cost per million tokens on an H100?

About $3.18 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 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.64.

What precision does Llama 4 Scout 17B-16E 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. The Meta repo is gated; geometry read from the unsloth mirror.