Meta Llama 3

Llama 3.3 70B Instruct hardware requirements

Fits on one Instinct MI350X OAM at BF16. For 1,000 tokens/s at 32k context you need 8 x H100 SXM5 80GB at $14.32/hr or 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $8.00/hr.

Parameters

70.6B

Active per token

All

dense model

KV cache per token

320 KB

16-bit cache

Native precision

BF16

Rent the cheapest fit: 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition from $1.00/hr Model card

Bare minimum

Fewest GPUs

1 x Instinct MI350X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 42.1
Per hour: $6.16
Per million tokens: $40.62
Memory used: 51%

Rent on DigitalOcean

Cheapest per hour

2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

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

Tokens/s: 17.9
Per hour: $1.00
Per million tokens: $15.50
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: 603
Per hour: $4.00
Per million tokens: $1.84
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; 21 more fit without a live price and appear in the table. Marker size grows with the tensor-parallel width. Roofline estimates.

GPU	TP	Memory used	Tokens/s	Per stream	Per hour	Per M tokens	kW
Instinct MI455X OAM	2	58%	~2.2k	68.5				Specs
Rubin SXM	2	87%	2.1k	64.7				Specs
Instinct MI355X OAM	2	87%	753	23.5			2.8	Specs
GB300 NVL72 GPU 288GB	2	87%	753	23.5	$17.24	$6.36		Rent
GB200 NVL72 GPU 186GB	4	68%	1.4k	44.6	$42.00	$8.18		Rent
Instinct MI350X OAM	2	87%	753	23.5	$12.32	$4.55	2.0	Rent
B300 SXM 262GB	2	96%	753	23.5	$14.80	$5.46	2.8	Rent
B200 SXM 180GB	4	70%	1.4k	42.9	$16.36	$3.31	4.0	Rent
B100 SXM 192GB	4	66%	1.4k	44.6			2.8	Specs
Instinct MI325X OAM	2	98%	564	17.6	$7.60	$3.74	2.0	Rent
Instinct MI300X 192GB	4	66%	945	29.5	$9.56	$2.81	3.0	Rent
BR100	8	100%	539	16.8			4.4	Specs
H200 SXM 141GB	4	90%	856	26.7	$11.96	$3.88	2.8	Rent
H100 SXM5 80GB	8	79%	1.1k	35.2	$14.32	$3.53	5.6	Rent
Data Center GPU Max 1550 128GB	4	99%	584	18.3			2.4	Specs
H200 NVL 141GB	4	90%	~856	26.7	$15.16	$4.92	2.4	Rent
H100 NVL 94GB	8	67%	~1.3k	41.0	$25.52	$5.40	3.2	Rent
H100 PCIe 80GB	8	79%	673	21.0	$20.00	$8.25	2.8	Rent
H800 PCIe 80GB	8	79%	673	21.0			2.8	Specs
RTX PRO 6000 Blackwell Workstation Edition	8	66%	603	18.9	$14.40	$6.63	4.8	Rent
RTX PRO 6000 Blackwell Server Edition	8	66%	~538	16.8	$4.72	$2.44	4.8	Rent
RTX PRO 6000 Blackwell Max-Q Workstation Edition	8	66%	603	18.9	$4.00	$1.84	2.4	Rent
Data Center GPU Max 1100 48GB 2 nodes	16	67%	730	22.8			4.8	Specs
MI250X 128GB	4	99%	584	18.3			2.2	Specs
RTX 6000 Ada 48GB 2 nodes	16	67%	~570	17.8	$13.44	$6.55	4.8	Rent
Instinct MI250 128GB	4	99%	584	18.3			2.0	Specs
L40S 48GB 2 nodes	16	67%	513	16.0	$13.92	$7.53	5.6	Rent
A800 PCIe 80GB	8	79%	652	20.4			2.4	Specs
A100 PCIe 80GB	8	79%	652	20.4	$10.80	$4.60	2.4	Rent
A100 SXM4 80GB	8	79%	687	21.5	$11.20	$4.53	3.2	Rent
A100 PCIe 40GB 2 nodes	16	81%	924	28.9			4.0	Specs
A100 SXM4 40GB 2 nodes	16	81%	924	28.9	$20.64	$6.21	6.4	Rent
RTX 5880 Ada 48GB 2 nodes	16	67%	~570	17.8			4.6	Specs
RTX PRO 5000 Blackwell 48GB 2 nodes	16	67%	~799	25.0	$15.36	$5.34	4.8	Rent
RTX PRO 5000 Blackwell 72GB	8	88%	~453	14.1			2.4	Specs
L40 48GB 2 nodes	16	67%	513	16.0	$13.12	$7.10	4.8	Rent
Instinct MI210 PCIe	8	100%	552	17.2			2.4	Specs
RTX A6000 48GB 2 nodes	16	67%	456	14.3	$8.00	$4.87	4.8	Rent
A40 48GB 2 nodes	16	67%	414	12.9	$7.84	$5.27	4.8	Rent
H20 141GB HBM3e	4	90%	713	22.3			1.6	Specs
H20 96GB	8	66%	1.3k	42.1			3.2	Specs
Radeon PRO W7900 Dual Slot 48GB 2 nodes	16	67%	513	16.0			4.7	Specs
Radeon PRO W7900 48GB 2 nodes	16	67%	513	16.0			4.7	Specs
L20 48GB 2 nodes	16	67%	513	16.0			4.4	Specs
Radeon PRO W7800 48GB 2 nodes	16	67%	513	16.0			4.2	Specs

AMD vs NVIDIA

At BF16 the NVIDIA pick is 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 603 tokens/s for $4.00/hr; the AMD pick is 4 x Instinct MI300X 192GB at about 945 tokens/s for $9.56/hr. Per rental dollar NVIDIA delivers 1.53x the tokens of AMD here.

NVIDIA 31 parts fit

Best value: 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 603 tok/s for $4.00/hr

Fewest GPUs: 2 x Rubin SXM, 2.1k tok/s

AMD 11 parts fit

Best value: 4 x Instinct MI300X 192GB, 945 tok/s for $9.56/hr

Fewest GPUs: 2 x Instinct MI455X OAM, 2.2k tok/s

Intel 2 parts fit

Fewest GPUs: 4 x Data Center GPU Max 1550 128GB, 584 tok/s

Other 1 parts fit

Fewest GPUs: 8 x BR100, 539 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	Replicas	Idle	Tokens/s	Per hour	Power	Per M tokens
A 1,000 x H200 SXM 141GB	250 x TP4	0	214k	$2,990	700 kW	$3.88
B 100 x GB200 NVL72 GPU 186GB	25 x TP4	0	36k	$1,050	kW	$8.18

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 MI350X 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 8 H100s across 1 replicas.

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

About $3.53 per million output tokens at BF16, 32,768 tokens of context and 32 streams, using the cheapest live on-demand price of $1.79 per GPU-hour. The cheapest part per token is 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.84.

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.