NVIDIA Llama Nemotron

Llama 3.1 Nemotron Ultra 253B hardware requirements

Needs at least 2 GPUs at BF16 (Instinct MI355X OAM). For 1,000 tokens/s at 32k context you need 64 x H100 SXM5 80GB at $117/hr or 112 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $56.00/hr.

Parameters

253B

Active per token

All

dense model

KV cache per token

256 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Llama 3.1 Nemotron Ultra 253B

The strongest part each vendor offers for Llama 3.1 Nemotron Ultra 253B 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.

4 x B300 SXM 262GB 4 x Instinct MI325X OAM 8 x Ascend 910C 8 x Gaudi 2 96GB 4 x TPU v7 192GB

Memory fills at: 4 x B300 SXM 262GB 58 users, 4 x Instinct MI325X OAM 55 users, 8 x Ascend 910C 55 users, 8 x Gaudi 2 96GB 26 users, 4 x TPU v7 192GB 26 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

2 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 22.1
Per hour: $11.98
Per million tokens: $150
Memory used: 92%

Rent on Daytona

Cheapest per hour

8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

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

Tokens/s: 17.7
Per hour: $4.00
Per million tokens: $62.85
Memory used: 69%

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: 152
Per hour: $8.00
Per million tokens: $14.59
Memory used: 53%

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	2	94 %	284 tok/s	42.5 tok/s	—	—	—	Specs →
2	Rubin SXM	4	70 %	443 tok/s	76.0 tok/s	—	—	—	Specs →
3	TPU 8t announced	4	94 %	~268 tok/s	22.5 tok/s	—	—	—	Specs →
4	TPU 8i announced	4	70 %	~251 tok/s	29.7 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	4	70 %	246 tok/s	27.6 tok/s	$23.96	$27.07	5.6 kW	Rent →
6	GB300 NVL72 GPU 288GB	4	70 %	245 tok/s	27.6 tok/s	$49.68	$56.40	—	Rent →
7	GB200 NVL72 GPU 186GB	8	55 %	462 tok/s	52.2 tok/s	$84.00	$50.49	—	Rent →
8	Instinct MI350X OAM	4	70 %	231 tok/s	27.6 tok/s	$24.64	$29.64	4.0 kW	Rent →
9	TPU v7 192GB	8	53 %	426 tok/s	48.2 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	2	94 %	~147 tok/s	42.5 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	4	77 %	226 tok/s	27.6 tok/s	$19.96	$24.48	5.6 kW	Rent →
12	B200 SXM 180GB	8	57 %	424 tok/s	50.2 tok/s	$32.72	$21.46	8.0 kW	Rent →
13	Ascend 970 announced	4	70 %	231 tok/s	49.7 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	8	53 %	353 tok/s	52.2 tok/s	—	—	5.6 kW	Specs →
15	Jalapeño announced	4	94 %	~203 tok/s	53.2 tok/s	—	—	2.8 kW	Specs →
16	Instinct MI325X OAM	4	79 %	140 tok/s	20.7 tok/s	$15.20	$30.24	4.0 kW	Rent →
17	Instinct MI300X 192GB	8	53 %	257 tok/s	34.6 tok/s	$19.12	$20.69	6.0 kW	Rent →
18	BR100 2 nodes	16	80 %	259 tok/s	18.4 tok/s	—	—	8.8 kW	Specs →
19	Ascend 960 announced	4	70 %	120 tok/s	33.2 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	8	72 %	202 tok/s	31.3 tok/s	$19.84	$27.28	5.6 kW	Rent →
21	H100 SXM5 80GB 2 nodes	16	64 %	328 tok/s	38.6 tok/s	$29.28	$24.81	11.2 kW	Rent →
22	Data Center GPU Max 1550 128GB	8	80 %	163 tok/s	21.4 tok/s	—	—	4.8 kW	Specs →
23	H200 NVL 141GB	8	72 %	~176 tok/s	31.3 tok/s	$28.91	$45.62	4.8 kW	Rent →
24	H100 NVL 94GB 2 nodes	16	54 %	~299 tok/s	44.9 tok/s	$51.04	$47.49	6.4 kW	Rent →
25	Ascend 910C	8	80 %	157 tok/s	20.9 tok/s	—	—	4.8 kW	Specs →
26	H100 PCIe 80GB 2 nodes	16	64 %	233 tok/s	23.0 tok/s	$40.00	$47.64	5.6 kW	Rent →
27	H800 PCIe 80GB 2 nodes	16	64 %	233 tok/s	23.0 tok/s	—	—	5.6 kW	Specs →
28	Ascend 950DT	8	71 %	120 tok/s	26.1 tok/s	—	—	—	Specs →
29	RTX PRO 6000 Blackwell Workstation Edition 2 nodes	16	53 %	169 tok/s	20.6 tok/s	$28.80	$47.33	9.6 kW	Rent →
30	Ascend 950PR	8	80 %	92.4 tok/s	10.4 tok/s	—	—	7.2 kW	Specs →
31	MTT S5000 2 nodes	16	64 %	~163 tok/s	18.4 tok/s	—	—	—	Specs →
32	RTX PRO 6000 Blackwell Server Edition 2 nodes	16	53 %	163 tok/s	18.4 tok/s	$11.04	$18.81	9.6 kW	Rent →
33	Gaudi 3 128GB	8	80 %	276 tok/s	24.1 tok/s	—	—	7.2 kW	Specs →
34	TPU v5p 95GB 2 nodes	16	54 %	172 tok/s	31.8 tok/s	—	—	7.2 kW	Specs →
35	RTX PRO 6000 Blackwell Max-Q Workstation Edition 2 nodes	16	53 %	152 tok/s	20.6 tok/s	$8.00	$14.59	4.8 kW	Rent →
36	Cloud AI 100 Ultra	8	80 %	~56.4 tok/s	3.6 tok/s	—	—	1.2 kW	Specs →
37	Gaudi 2 96GB 2 nodes	16	53 %	160 tok/s	28.2 tok/s	$17.60	$30.51	9.6 kW	Rent →
38	Atlas 350	8	91 %	79.2 tok/s	9.1 tok/s	—	—	4.8 kW	Specs →
39	Ascend 910B 2 nodes	16	80 %	138 tok/s	18.4 tok/s	—	—	6.4 kW	Specs →
40	MI250X 128GB	8	80 %	85.6 tok/s	21.4 tok/s	—	—	4.5 kW	Specs →
41	Instinct MI250 128GB	8	80 %	81.5 tok/s	21.4 tok/s	—	—	4.0 kW	Specs →
42	A800 PCIe 80GB 2 nodes	16	64 %	118 tok/s	22.3 tok/s	—	—	4.8 kW	Specs →
43	A800 SXM4 2 nodes	16	64 %	119 tok/s	23.5 tok/s	—	—	6.4 kW	Specs →
44	A100 PCIe 80GB 2 nodes	16	64 %	118 tok/s	22.3 tok/s	$21.60	$51.06	4.8 kW	Rent →
45	A100 SXM4 80GB 2 nodes	16	64 %	119 tok/s	23.5 tok/s	$22.40	$52.50	6.4 kW	Rent →
46	RTX PRO 5000 Blackwell 72GB 2 nodes	16	71 %	~94.2 tok/s	15.5 tok/s	—	—	4.8 kW	Specs →
47	Instinct MI210 PCIe 2 nodes	16	80 %	71.9 tok/s	18.9 tok/s	—	—	4.8 kW	Specs →
48	H20 141GB HBM3e	8	72 %	36.2 tok/s	26.1 tok/s	—	—	3.2 kW	Specs →
49	H20 96GB 2 nodes	16	53 %	63.9 tok/s	46.1 tok/s	—	—	6.4 kW	Specs →
50	Atlas 300I Duo 96GB 2 nodes	16	53 %	44.5 tok/s	4.7 tok/s	—	—	2.4 kW	Specs →
51	ATOM-Max 2 nodes	16	80 %	50.1 tok/s	11.8 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 152 tokens/s for $8.00/hr; the AMD pick is 8 x Instinct MI300X 192GB at about 257 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, 152 tok/s for $8.00/hr

Fewest GPUs: 4 x Rubin SXM, 443 tok/s

AMD 8 parts fit

Best value: 8 x Instinct MI300X 192GB, 257 tok/s for $19.12/hr

Fewest GPUs: 2 x Instinct MI455X, 284 tok/s

Huawei 6 parts fit

Fewest GPUs: 8 x Ascend 910C, 157 tok/s

Intel 3 parts fit

Best value: 16 x Gaudi 2 96GB, 160 tok/s for $17.60/hr

Fewest GPUs: 8 x Gaudi 3 128GB, 276 tok/s

Google 2 parts fit

Fewest GPUs: 8 x TPU v7 192GB, 426 tok/s

Other 4 parts fit

Fewest GPUs: 8 x Cloud AI 100 Ultra, 56.4 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	25k tok/s	$2,480	700 kW	$27.28
B	104 x GB200 NVL72 GPU 186GB	13 x TP8	0	6.0k tok/s	$1,092	—	$50.49

Size Llama 3.1 Nemotron Ultra 253B 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 Nemotron Ultra 253B need?

The weights take 507 GB at the native BF16 precision (507 GB at BF16, 253 GB at FP8, 143 GB at INT4). Each concurrent stream adds 256 KB of KV cache per token: 1.1 GB at 4,096 tokens and 8.6 GB at 32,768 tokens.

What is the bare minimum to run Llama 3.1 Nemotron Ultra 253B?

2 x Instinct MI355X OAM at BF16 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $4.00 per hour.

How many H100s do you need to run Llama 3.1 Nemotron Ultra 253B?

8 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 64 H100s across 4 replicas.

How much does Llama 3.1 Nemotron Ultra 253B cost per million tokens on an H100?

About $24.81 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 $14.59.

What precision does Llama 3.1 Nemotron Ultra 253B 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. Llama 3.1 405B pruned by neural architecture search: 64 of 162 layers keep attention (256 KB per token). NVIDIA's card serves it with tensor parallelism across 8 GPUs.