Meta Llama 3

Llama 3.1 8B Instruct hardware requirements

Fits on one Instinct MI355X OAM at BF16. For 1,000 tokens/s at 32k context you need 4 x H100 SXM5 80GB at $7.32/hr or 6 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $3.00/hr.

Parameters

8.0B

Active per token

All

dense model

KV cache per token

128 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Llama 3.1 8B Instruct

The strongest part each vendor offers for Llama 3.1 8B Instruct 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 1 x Inferentia2

Memory fills at: 1 x GB300 NVL72 GPU 288GB 60 users, 1 x Instinct MI355X OAM 60 users, 1 x Ascend 910C 24 users, 1 x Gaudi 2 96GB 17 users, 1 x TPU v7 192GB 39 users, 1 x Inferentia2 3 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 357
Per hour: $5.99
Per million tokens: $4.66
Memory used: 6%

Rent on Daytona

Cheapest per hour

1 x RTX A5000 24GB

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

Tokens/s: 31.1
Per hour: $0.27
Per million tokens: $2.41
Memory used: 75%

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: 346
Per hour: $1.00
Per million tokens: $0.80
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	37 %	2.3k tok/s	114 tok/s	—	—	—	Specs →
2	Rubin SXM	1	55 %	2.0k tok/s	107 tok/s	—	—	—	Specs →
3	TPU 8t announced	1	74 %	~802 tok/s	31.9 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	55 %	~928 tok/s	42.0 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	55 %	881 tok/s	39.1 tok/s	$5.99	$1.89	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	55 %	880 tok/s	39.1 tok/s	$12.42	$3.92	—	Rent →
7	GB200 NVL72 GPU 186GB	1	86 %	880 tok/s	39.1 tok/s	$10.50	$3.32	—	Rent →
8	Instinct MI350X OAM	1	55 %	859 tok/s	39.1 tok/s	$6.16	$1.99	1.0 kW	Rent →
9	TPU v7 192GB	1	83 %	812 tok/s	36.1 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	37 %	~1.6k tok/s	114 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	61 %	852 tok/s	39.1 tok/s	$4.99	$1.63	1.4 kW	Rent →
12	B200 SXM 180GB	1	89 %	830 tok/s	37.6 tok/s	$4.09	$1.37	1.0 kW	Rent →
13	Ascend 970 announced	1	55 %	1.2k tok/s	70.4 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	1	83 %	780 tok/s	39.1 tok/s	—	—	0.7 kW	Specs →
15	Jalapeño announced	1	74 %	~1.1k tok/s	75.2 tok/s	—	—	0.7 kW	Specs →
16	Instinct MI325X OAM	1	62 %	584 tok/s	29.3 tok/s	$3.80	$1.81	1.0 kW	Rent →
17	Instinct MI300X 192GB	1	83 %	540 tok/s	25.9 tok/s	$2.39	$1.23	0.8 kW	Rent →
18	BR100	4	63 %	747 tok/s	28.1 tok/s	—	—	2.2 kW	Specs →
19	Ascend 960 announced	1	55 %	663 tok/s	46.9 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	2	57 %	870 tok/s	44.6 tok/s	$4.96	$1.58	1.4 kW	Rent →
21	H100 SXM5 80GB	4	50 %	1.3k tok/s	58.9 tok/s	$7.32	$1.56	2.8 kW	Rent →
22	TPU v6e 32GB	8	64 %	1.4k tok/s	54.4 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	2	63 %	643 tok/s	30.4 tok/s	—	—	1.2 kW	Specs →
24	H200 NVL 141GB	2	57 %	~811 tok/s	44.6 tok/s	$7.23	$2.47	1.2 kW	Rent →
25	H100 NVL 94GB	2	86 %	~717 tok/s	36.2 tok/s	$6.38	$2.47	0.8 kW	Rent →
26	Ascend 910C	2	63 %	622 tok/s	29.7 tok/s	—	—	1.2 kW	Specs →
27	H100 PCIe 80GB	4	50 %	835 tok/s	35.2 tok/s	$10.00	$3.33	1.4 kW	Rent →
28	H800 PCIe 80GB	4	50 %	835 tok/s	35.2 tok/s	—	—	1.4 kW	Specs →
29	Ascend 950DT	2	56 %	605 tok/s	37.1 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	2	84 %	362 tok/s	16.6 tok/s	$3.60	$2.76	1.2 kW	Rent →
31	Ascend 950PR	2	63 %	334 tok/s	14.9 tok/s	—	—	1.8 kW	Specs →
32	MTT S5000	4	50 %	~633 tok/s	28.1 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	2	84 %	334 tok/s	14.8 tok/s	$1.38	$1.15	1.2 kW	Rent →
34	Gaudi 3 128GB	2	63 %	852 tok/s	34.3 tok/s	—	—	1.8 kW	Specs →
35	TPU v5p 95GB	2	85 %	458 tok/s	25.7 tok/s	—	—	0.9 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	2	84 %	346 tok/s	16.6 tok/s	$1.00	$0.80	0.6 kW	Rent →
37	Cloud AI 100 Ultra	2	63 %	~140 tok/s	5.1 tok/s	—	—	0.3 kW	Specs →
38	Gaudi 2 96GB	2	84 %	416 tok/s	22.7 tok/s	$2.20	$1.47	1.2 kW	Rent →
39	Atlas 350	2	72 %	290 tok/s	13.0 tok/s	—	—	1.2 kW	Specs →
40	Data Center GPU Max 1100 48GB	4	85 %	499 tok/s	21.6 tok/s	—	—	1.2 kW	Specs →
41	GeForce RTX 5090 32GB	8	64 %	895 tok/s	59.5 tok/s	$9.52	$2.95	4.6 kW	Rent →
42	Ascend 910B	4	63 %	590 tok/s	28.1 tok/s	—	—	1.6 kW	Specs →
43	MI250X 128GB	2	63 %	457 tok/s	30.4 tok/s	—	—	1.1 kW	Specs →
44	RTX 6000 Ada 48GB	4	85 %	~401 tok/s	16.9 tok/s	$3.88	$2.69	1.2 kW	Rent →
45	Instinct MI250 128GB	2	63 %	444 tok/s	30.4 tok/s	—	—	1.0 kW	Specs →
46	L40S 48GB	4	85 %	370 tok/s	15.2 tok/s	$3.04	$2.28	1.4 kW	Rent →
47	A800 PCIe 80GB	4	50 %	599 tok/s	34.0 tok/s	—	—	1.2 kW	Specs →
48	A800 SXM4	4	50 %	617 tok/s	35.9 tok/s	—	—	1.6 kW	Specs →
49	A100 PCIe 80GB	4	50 %	599 tok/s	34.0 tok/s	$5.40	$2.50	1.2 kW	Rent →
50	A100 SXM4 80GB	4	50 %	617 tok/s	35.9 tok/s	$5.60	$2.52	1.6 kW	Rent →
51	A100 PCIe 40GB	8	51 %	998 tok/s	51.7 tok/s	—	—	2.0 kW	Specs →
52	A100 SXM4 40GB	8	51 %	998 tok/s	51.7 tok/s	$12.49	$3.48	3.2 kW	Rent →
53	GeForce RTX 5090 D V2 24GB	8	87 %	~895 tok/s	44.7 tok/s	—	—	4.6 kW	Specs →
54	RTX 5880 Ada 48GB	4	85 %	~371 tok/s	16.9 tok/s	—	—	1.1 kW	Specs →
55	TPU v4 32GB	8	64 %	810 tok/s	39.9 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB	8	64 %	~474 tok/s	19.1 tok/s	$6.64	$3.89	2.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	4	85 %	~448 tok/s	23.6 tok/s	$3.84	$2.38	1.2 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	4	56 %	~448 tok/s	23.6 tok/s	—	—	1.2 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition	8	64 %	560 tok/s	26.6 tok/s	—	—	1.3 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB	8	64 %	602 tok/s	29.8 tok/s	$5.76	$2.66	1.6 kW	Rent →
61	Radeon AI PRO R9700 32GB	8	64 %	472 tok/s	21.3 tok/s	—	—	2.4 kW	Specs →
62	Inferentia2	8	64 %	556 tok/s	27.2 tok/s	—	—	—	Specs →
63	Trainium	8	64 %	556 tok/s	27.2 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB	8	64 %	474 tok/s	40.8 tok/s	—	—	2.4 kW	Specs →
65	Arc Pro B70 32GB	8	64 %	~450 tok/s	20.2 tok/s	—	—	1.8 kW	Specs →
66	L40 48GB	4	85 %	299 tok/s	15.2 tok/s	$3.28	$3.05	1.2 kW	Rent →
67	Instinct MI210 PCIe	4	63 %	420 tok/s	28.8 tok/s	—	—	1.2 kW	Specs →
68	GeForce RTX 4090 24GB	8	87 %	594 tok/s	33.5 tok/s	$7.12	$3.33	3.6 kW	Rent →
69	A30 24GB	8	87 %	568 tok/s	31.0 tok/s	—	—	1.3 kW	Specs →
70	RTX 4500 Ada 24GB	8	87 %	~338 tok/s	14.4 tok/s	—	—	1.7 kW	Specs →
71	RTX A6000 48GB	4	85 %	261 tok/s	13.5 tok/s	$2.00	$2.13	1.2 kW	Rent →
72	A40 48GB	4	85 %	243 tok/s	12.2 tok/s	$2.36	$2.70	1.2 kW	Rent →
73	H20 141GB HBM3e	2	57 %	260 tok/s	37.1 tok/s	—	—	0.8 kW	Specs →
74	H20 96GB	2	84 %	260 tok/s	37.1 tok/s	—	—	0.8 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB	8	87 %	~446 tok/s	22.3 tok/s	$4.56	$2.84	1.2 kW	Rent →
76	BI-V100 Tiangai 100	8	64 %	615 tok/s	39.9 tok/s	—	—	2.0 kW	Specs →
77	Atlas 300I Duo 48GB	4	85 %	166 tok/s	7.2 tok/s	—	—	0.6 kW	Specs →
78	Atlas 300I Duo 96GB	2	84 %	87.6 tok/s	3.8 tok/s	—	—	0.3 kW	Specs →
79	RTX A5500 24GB	8	87 %	~469 tok/s	25.5 tok/s	—	—	1.8 kW	Specs →
80	Wormhole n300s	8	87 %	388 tok/s	19.1 tok/s	—	—	2.4 kW	Specs →
81	V100S PCIe 32GB	8	64 %	561 tok/s	37.7 tok/s	$7.04	$3.49	2.0 kW	Rent →
82	ATOM-Max	4	63 %	281 tok/s	18.0 tok/s	—	—	1.4 kW	Specs →
83	A10 24GB	8	87 %	391 tok/s	19.9 tok/s	$8.81	$6.27	1.2 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	4	85 %	252 tok/s	15.2 tok/s	—	—	1.2 kW	Specs →
85	Radeon PRO W7900 48GB	4	85 %	252 tok/s	15.2 tok/s	—	—	1.2 kW	Specs →
86	Radeon RX 7900 XTX 24GB	8	87 %	503 tok/s	31.9 tok/s	—	—	2.8 kW	Specs →
87	L4 24GB	8	87 %	240 tok/s	10.0 tok/s	$4.72	$5.45	0.6 kW	Rent →
88	L20 48GB	4	85 %	249 tok/s	15.2 tok/s	—	—	1.1 kW	Specs →
89	Radeon AI PRO R9600D 32GB	8	64 %	367 tok/s	21.3 tok/s	—	—	1.2 kW	Specs →
90	Arc Pro B65 32GB	8	64 %	~356 tok/s	20.2 tok/s	—	—	1.6 kW	Specs →
91	Arc Pro B60 24GB	8	87 %	~301 tok/s	15.2 tok/s	—	—	1.6 kW	Specs →
92	L2 24GB	8	87 %	226 tok/s	10.0 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB	8	87 %	~289 tok/s	14.4 tok/s	—	—	0.6 kW	Specs →
94	MLU370-X8	4	85 %	188 tok/s	10.8 tok/s	—	—	1.0 kW	Specs →
95	Radeon PRO W7800 48GB	4	85 %	215 tok/s	15.2 tok/s	—	—	1.0 kW	Specs →
96	Radeon PRO W7800 32GB	8	64 %	333 tok/s	19.1 tok/s	—	—	2.1 kW	Specs →
97	GeForce RTX 3090 24GB	8	87 %	363 tok/s	31.1 tok/s	$4.00	$3.06	2.8 kW	Rent →
98	A10G 24GB	8	87 %	299 tok/s	19.9 tok/s	$8.05	$7.46	2.4 kW	Rent →
99	Atlas 300V Pro	4	85 %	83.2 tok/s	3.6 tok/s	—	—	—	Specs →
100	Atlas 300I Pro	8	87 %	157 tok/s	6.8 tok/s	—	—	—	Specs →
101	Atlas 300V	8	87 %	141 tok/s	6.8 tok/s	—	—	0.6 kW	Specs →
102	GeForce RTX 3090 Ti 24GB	8	87 %	248 tok/s	33.5 tok/s	$3.68	$4.12	3.6 kW	Rent →
103	Radeon PRO W6800 32GB	8	64 %	188 tok/s	17.0 tok/s	—	—	2.0 kW	Specs →
104	RTX A5000 24GB	8	87 %	176 tok/s	25.5 tok/s	$2.16	$3.41	1.8 kW	Rent →
105	Radeon PRO V710 28GB	8	74 %	152 tok/s	14.9 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 346 tokens/s for $1.00/hr; the AMD pick is 1 x Instinct MI300X 192GB at about 540 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, 346 tok/s for $1.00/hr

Fewest GPUs: 1 x Rubin SXM, 2.0k tok/s

AMD 18 parts fit

Best value: 1 x Instinct MI300X 192GB, 540 tok/s for $2.39/hr

Fewest GPUs: 1 x Instinct MI455X, 2.3k tok/s

Huawei 10 parts fit

Fewest GPUs: 2 x Ascend 910C, 622 tok/s

Intel 7 parts fit

Best value: 2 x Gaudi 2 96GB, 416 tok/s for $2.20/hr

Fewest GPUs: 2 x Gaudi 3 128GB, 852 tok/s

Google 4 parts fit

Fewest GPUs: 1 x TPU v7 192GB, 812 tok/s

AWS 2 parts fit

Fewest GPUs: 8 x Inferentia2, 556 tok/s

Other 7 parts fit

Fewest GPUs: 2 x Cloud AI 100 Ultra, 140 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	435k tok/s	$2,480	700 kW	$1.58
B	100 x GB200 NVL72 GPU 186GB	100 x TP1	0	88k tok/s	$1,050	—	$3.32

Size Llama 3.1 8B Instruct 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 3.1 8B Instruct need?

The weights take 16.1 GB at the native BF16 precision (16.1 GB at BF16, 8.0 GB at FP8, 4.5 GB at INT4). Each concurrent stream adds 128 KB of KV cache per token: 0.5 GB at 4,096 tokens and 4.3 GB at 32,768 tokens.

What is the bare minimum to run Llama 3.1 8B Instruct?

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 A5000 24GB at $0.27 per hour.

How many H100s do you need to run Llama 3.1 8B Instruct?

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 4 H100s across 1 replicas.

How much does Llama 3.1 8B Instruct cost per million tokens on an H100?

About $1.56 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 $0.80.

What precision does Llama 3.1 8B Instruct 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. The Meta repo is gated; geometry read from the unsloth mirror, which ships the same config.