Meta Llama 3

Llama 3.3 70B Instruct hardware requirements

Fits on one Instinct MI355X OAM at BF16. For 1,000 tokens/s at 32k context you need 16 x H100 SXM5 80GB at $29.28/hr or 40 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $20.00/hr.

Parameters

70.6B

Active per token

All

dense model

KV cache per token

320 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Llama 3.3 70B Instruct

The strongest part each vendor offers for Llama 3.3 70B 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 2 x Ascend 910C 2 x Gaudi 3 128GB 1 x TPU v7 192GB 8 x Inferentia2

1 x GB300 NVL72 GPU 288GB and 1 x Instinct MI355X OAM draw the same line: identical memory and bandwidth give identical curves.

Memory fills at: 1 x GB300 NVL72 GPU 288GB 12 users, 1 x Instinct MI355X OAM 12 users, 2 x Ascend 910C 9 users, 2 x Gaudi 3 128GB 9 users, 1 x TPU v7 192GB 4 users, 8 x Inferentia2 9 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 41.5
Per hour: $5.99
Per million tokens: $40.05
Memory used: 51%

Rent on Daytona

Cheapest per hour

2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

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

Tokens/s: 17.6
Per hour: $1.00
Per million tokens: $15.77
Memory used: 78%

Rent on RunPod

Cheapest per million tokens

8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 242
Per hour: $4.00
Per million tokens: $4.60
Memory used: 66%

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; 43 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	2	58 %	812 tok/s	68.5 tok/s	—	—	—	Specs →
2	Rubin SXM	2	87 %	686 tok/s	64.7 tok/s	—	—	—	Specs →
3	TPU 8t announced	4	58 %	~661 tok/s	36.4 tok/s	—	—	—	Specs →
4	TPU 8i announced	2	87 %	~360 tok/s	25.3 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	2	87 %	347 tok/s	23.5 tok/s	$11.98	$9.58	2.8 kW	Rent →
6	GB300 NVL72 GPU 288GB	2	87 %	346 tok/s	23.5 tok/s	$24.84	$19.93	—	Rent →
7	GB200 NVL72 GPU 186GB	4	68 %	656 tok/s	44.6 tok/s	$42.00	$17.78	—	Rent →
8	Instinct MI350X OAM	2	87 %	331 tok/s	23.5 tok/s	$12.32	$10.32	2.0 kW	Rent →
9	TPU v7 192GB	4	66 %	605 tok/s	41.1 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	2	58 %	~465 tok/s	68.5 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	2	96 %	327 tok/s	23.5 tok/s	$9.98	$8.49	2.8 kW	Rent →
12	B200 SXM 180GB	4	70 %	609 tok/s	42.9 tok/s	$16.36	$7.47	4.0 kW	Rent →
13	Ascend 970 announced	2	87 %	372 tok/s	42.3 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	4	66 %	533 tok/s	44.6 tok/s	—	—	2.8 kW	Specs →
15	Jalapeño announced	4	58 %	~635 tok/s	85.8 tok/s	—	—	2.8 kW	Specs →
16	Instinct MI325X OAM	2	98 %	210 tok/s	17.6 tok/s	$7.60	$10.04	2.0 kW	Rent →
17	Instinct MI300X 192GB	4	66 %	380 tok/s	29.5 tok/s	$9.56	$6.99	3.0 kW	Rent →
18	BR100	8	100 %	342 tok/s	16.8 tok/s	—	—	4.4 kW	Specs →
19	Ascend 960 announced	2	87 %	200 tok/s	28.2 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	4	90 %	308 tok/s	26.7 tok/s	$9.92	$8.95	2.8 kW	Rent →
21	H100 SXM5 80GB	8	79 %	503 tok/s	35.2 tok/s	$14.64	$8.08	5.6 kW	Rent →
22	Data Center GPU Max 1550 128GB	4	99 %	240 tok/s	18.3 tok/s	—	—	2.4 kW	Specs →
23	H200 NVL 141GB	4	90 %	~275 tok/s	26.7 tok/s	$14.46	$14.58	2.4 kW	Rent →
24	H100 NVL 94GB	8	67 %	~484 tok/s	41.0 tok/s	$25.52	$14.64	3.2 kW	Rent →
25	Ascend 910C	4	99 %	231 tok/s	17.8 tok/s	—	—	2.4 kW	Specs →
26	H100 PCIe 80GB	8	79 %	342 tok/s	21.0 tok/s	$20.00	$16.25	2.8 kW	Rent →
27	H800 PCIe 80GB	8	79 %	342 tok/s	21.0 tok/s	—	—	2.8 kW	Specs →
28	Ascend 950DT	4	88 %	194 tok/s	22.3 tok/s	—	—	—	Specs →
29	RTX PRO 6000 Blackwell Workstation Edition	8	66 %	262 tok/s	18.9 tok/s	$14.40	$15.28	4.8 kW	Rent →
30	Ascend 950PR	4	99 %	131 tok/s	8.9 tok/s	—	—	3.6 kW	Specs →
31	MTT S5000	8	79 %	~248 tok/s	16.8 tok/s	—	—	—	Specs →
32	RTX PRO 6000 Blackwell Server Edition	8	66 %	248 tok/s	16.8 tok/s	$5.52	$6.19	4.8 kW	Rent →
33	Gaudi 3 128GB	4	99 %	365 tok/s	20.6 tok/s	—	—	3.6 kW	Specs →
34	TPU v5p 95GB	8	67 %	290 tok/s	29.1 tok/s	—	—	3.6 kW	Specs →
35	RTX PRO 6000 Blackwell Max-Q Workstation Edition	8	66 %	242 tok/s	18.9 tok/s	$4.00	$4.60	2.4 kW	Rent →
36	Cloud AI 100 Ultra	4	99 %	~66.8 tok/s	3.1 tok/s	—	—	0.6 kW	Specs →
37	Gaudi 2 96GB	8	66 %	268 tok/s	25.8 tok/s	—	—	4.8 kW	Specs →
38	Atlas 350	8	57 %	213 tok/s	14.7 tok/s	—	—	4.8 kW	Specs →
39	Data Center GPU Max 1100 48GB 2 nodes	16	67 %	352 tok/s	22.8 tok/s	—	—	4.8 kW	Specs →
40	Ascend 910B	8	100 %	218 tok/s	16.8 tok/s	—	—	3.2 kW	Specs →
41	MI250X 128GB	4	99 %	141 tok/s	18.3 tok/s	—	—	2.2 kW	Specs →
42	RTX 6000 Ada 48GB 2 nodes	16	67 %	~290 tok/s	17.8 tok/s	$15.84	$15.17	4.8 kW	Rent →
43	Instinct MI250 128GB	4	99 %	135 tok/s	18.3 tok/s	—	—	2.0 kW	Specs →
44	L40S 48GB 2 nodes	16	67 %	274 tok/s	16.0 tok/s	$12.16	$12.34	5.6 kW	Rent →
45	A800 PCIe 80GB	8	79 %	199 tok/s	20.4 tok/s	—	—	2.4 kW	Specs →
46	A800 SXM4	8	79 %	202 tok/s	21.5 tok/s	—	—	3.2 kW	Specs →
47	A100 PCIe 80GB	8	79 %	199 tok/s	20.4 tok/s	$10.80	$15.08	2.4 kW	Rent →
48	A100 SXM4 80GB	8	79 %	202 tok/s	21.5 tok/s	$10.80	$14.85	3.2 kW	Rent →
49	A100 PCIe 40GB 2 nodes	16	81 %	327 tok/s	28.9 tok/s	—	—	4.0 kW	Specs →
50	A100 SXM4 40GB 2 nodes	16	81 %	327 tok/s	28.9 tok/s	$24.98	$21.24	6.4 kW	Rent →
51	RTX 5880 Ada 48GB 2 nodes	16	67 %	~251 tok/s	17.8 tok/s	—	—	4.6 kW	Specs →
52	RTX PRO 5000 Blackwell 48GB 2 nodes	16	67 %	~274 tok/s	25.0 tok/s	$15.36	$15.55	4.8 kW	Rent →
53	RTX PRO 5000 Blackwell 72GB	8	88 %	~155 tok/s	14.1 tok/s	—	—	2.4 kW	Specs →
54	L40 48GB 2 nodes	16	67 %	187 tok/s	16.0 tok/s	$13.12	$19.53	4.8 kW	Rent →
55	Instinct MI210 PCIe	8	100 %	128 tok/s	17.2 tok/s	—	—	2.4 kW	Specs →
56	RTX A6000 48GB 2 nodes	16	67 %	162 tok/s	14.3 tok/s	$8.00	$13.73	4.8 kW	Rent →
57	A40 48GB 2 nodes	16	67 %	153 tok/s	12.9 tok/s	$9.44	$17.15	4.8 kW	Rent →
58	H20 141GB HBM3e	4	90 %	65.3 tok/s	22.3 tok/s	—	—	1.6 kW	Specs →
59	H20 96GB	8	66 %	123 tok/s	42.1 tok/s	—	—	3.2 kW	Specs →
60	Atlas 300I Duo 48GB 2 nodes	16	67 %	117 tok/s	7.6 tok/s	—	—	2.4 kW	Specs →
61	Atlas 300I Duo 96GB	8	66 %	66.4 tok/s	4.3 tok/s	—	—	1.2 kW	Specs →
62	ATOM-Max	8	100 %	87.6 tok/s	10.8 tok/s	—	—	2.8 kW	Specs →
63	Radeon PRO W7900 Dual Slot 48GB 2 nodes	16	67 %	144 tok/s	16.0 tok/s	—	—	4.7 kW	Specs →
64	Radeon PRO W7900 48GB 2 nodes	16	67 %	144 tok/s	16.0 tok/s	—	—	4.7 kW	Specs →
65	L20 48GB 2 nodes	16	67 %	141 tok/s	16.0 tok/s	—	—	4.4 kW	Specs →
66	MLU370-X8 2 nodes	16	67 %	109 tok/s	11.4 tok/s	—	—	4.0 kW	Specs →
67	Radeon PRO W7800 48GB 2 nodes	16	67 %	114 tok/s	16.0 tok/s	—	—	4.2 kW	Specs →
68	Atlas 300V Pro 2 nodes	16	67 %	58.7 tok/s	3.8 tok/s	—	—	—	Specs →

By vendor

At BF16 the NVIDIA pick is 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 242 tokens/s for $4.00/hr; the AMD pick is 4 x Instinct MI300X 192GB at about 380 tokens/s for $9.56/hr. Per rental dollar NVIDIA delivers 1.52x the tokens of AMD here.

NVIDIA 32 parts fit

Best value: 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 242 tok/s for $4.00/hr

Fewest GPUs: 2 x Rubin SXM, 686 tok/s

AMD 11 parts fit

Best value: 4 x Instinct MI300X 192GB, 380 tok/s for $9.56/hr

Fewest GPUs: 2 x Instinct MI455X, 812 tok/s

Huawei 8 parts fit

Fewest GPUs: 4 x Ascend 910C, 231 tok/s

Intel 4 parts fit

Fewest GPUs: 4 x Gaudi 3 128GB, 365 tok/s

Google 2 parts fit

Fewest GPUs: 4 x TPU v7 192GB, 605 tok/s

Other 5 parts fit

Fewest GPUs: 4 x Cloud AI 100 Ultra, 66.8 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	77k tok/s	$2,480	700 kW	$8.95
B	100 x GB200 NVL72 GPU 186GB	25 x TP4	0	16k tok/s	$1,050	—	$17.78

Size Llama 3.3 70B 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 3.3 70B Instruct need?

The weights take 141 GB at the native BF16 precision (141 GB at BF16, 70.6 GB at FP8, 39.7 GB at INT4). Each concurrent stream adds 320 KB of KV cache per token: 1.3 GB at 4,096 tokens and 10.7 GB at 32,768 tokens.

What is the bare minimum to run Llama 3.3 70B 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 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.00 per hour.

How many H100s do you need to run Llama 3.3 70B Instruct?

2 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 16 H100s across 2 replicas.

How much does Llama 3.3 70B Instruct cost per million tokens on an H100?

About $8.08 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 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $4.60.

What precision does Llama 3.3 70B 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.