NVIDIA Nemotron 3

Nemotron 3 Nano 4B hardware requirements

Fits on one Instinct MI355X OAM at BF16. For 1,000 tokens/s at 32k context you need 1 x H100 SXM5 80GB at $1.83/hr or 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.00/hr.

Parameters

4.0B

Active per token

All

dense model

KV cache per token

16 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Nemotron 3 Nano 4B

The strongest part each vendor offers for Nemotron 3 Nano 4B 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 433 users, 1 x Instinct MI355X OAM 433 users, 1 x Ascend 910C 184 users, 1 x Gaudi 2 96GB 134 users, 1 x TPU v7 192GB 284 users, 1 x Inferentia2 35 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 731
Per hour: $5.99
Per million tokens: $2.28
Memory used: 3%

Rent on Daytona

Cheapest per hour

1 x RTX A5000 24GB

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

Tokens/s: 63.6
Per hour: $0.27
Per million tokens: $1.18
Memory used: 37%

Rent on RunPod

Cheapest per million tokens

1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 626
Per hour: $0.50
Per million tokens: $0.22
Memory used: 30%

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	7 %	7.5k tok/s	628 tok/s	—	—	—	Specs →
2	Rubin SXM	1	10 %	6.4k tok/s	593 tok/s	—	—	—	Specs →
3	TPU 8t announced	1	13 %	~3.2k tok/s	176 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	10 %	~3.3k tok/s	232 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	10 %	3.2k tok/s	216 tok/s	$5.99	$0.52	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	10 %	3.2k tok/s	216 tok/s	$12.42	$1.08	—	Rent →
7	GB200 NVL72 GPU 186GB	1	16 %	3.2k tok/s	216 tok/s	$10.50	$0.91	—	Rent →
8	Instinct MI350X OAM	1	10 %	3.1k tok/s	216 tok/s	$6.16	$0.56	1.0 kW	Rent →
9	TPU v7 192GB	1	15 %	3.0k tok/s	199 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	7 %	~4.3k tok/s	628 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	11 %	3.0k tok/s	216 tok/s	$4.99	$0.46	1.4 kW	Rent →
12	B200 SXM 180GB	1	16 %	3.0k tok/s	208 tok/s	$4.09	$0.38	1.0 kW	Rent →
13	Ascend 970 announced	1	10 %	3.5k tok/s	388 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	1	15 %	2.6k tok/s	216 tok/s	—	—	0.7 kW	Specs →
15	Jalapeño announced	1	13 %	~3.1k tok/s	415 tok/s	—	—	0.7 kW	Specs →
16	Instinct MI325X OAM	1	11 %	2.0k tok/s	162 tok/s	$3.80	$0.54	1.0 kW	Rent →
17	Instinct MI300X 192GB	1	15 %	1.9k tok/s	143 tok/s	$2.39	$0.36	0.8 kW	Rent →
18	BR100	1	46 %	884 tok/s	43.2 tok/s	—	—	0.6 kW	Specs →
19	Ascend 960 announced	1	10 %	1.9k tok/s	259 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	1	21 %	1.5k tok/s	129 tok/s	$2.48	$0.46	0.7 kW	Rent →
21	H100 SXM5 80GB	1	36 %	1.3k tok/s	90.3 tok/s	$1.83	$0.39	0.7 kW	Rent →
22	TPU v6e 32GB	1	93 %	861 tok/s	44.2 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	1	23 %	1.2k tok/s	88.4 tok/s	—	—	0.6 kW	Specs →
24	H200 NVL 141GB	1	21 %	~1.3k tok/s	129 tok/s	$3.61	$0.74	0.6 kW	Rent →
25	H100 NVL 94GB	1	31 %	~1.3k tok/s	105 tok/s	$3.19	$0.71	0.4 kW	Rent →
26	Ascend 910C	1	23 %	1.1k tok/s	86.3 tok/s	—	—	0.6 kW	Specs →
27	H100 PCIe 80GB	1	36 %	884 tok/s	53.9 tok/s	$2.50	$0.79	0.3 kW	Rent →
28	H800 PCIe 80GB	1	36 %	884 tok/s	53.9 tok/s	—	—	0.3 kW	Specs →
29	Ascend 950DT	1	20 %	950 tok/s	108 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	1	30 %	678 tok/s	48.3 tok/s	$1.80	$0.74	0.6 kW	Rent →
31	Ascend 950PR	1	23 %	642 tok/s	43.2 tok/s	—	—	0.9 kW	Specs →
32	MTT S5000	1	36 %	~642 tok/s	43.2 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	1	30 %	641 tok/s	43.1 tok/s	$0.69	$0.30	0.6 kW	Rent →
34	Gaudi 3 128GB	1	23 %	1.8k tok/s	99.8 tok/s	—	—	0.9 kW	Specs →
35	TPU v5p 95GB	1	31 %	753 tok/s	74.6 tok/s	—	—	0.5 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	1	30 %	626 tok/s	48.3 tok/s	$0.50	$0.22	0.3 kW	Rent →
37	Cloud AI 100 Ultra	1	23 %	~325 tok/s	14.8 tok/s	—	—	0.1 kW	Specs →
38	Gaudi 2 96GB	1	30 %	695 tok/s	66.1 tok/s	$1.10	$0.44	0.6 kW	Rent →
39	Atlas 350	1	26 %	553 tok/s	37.8 tok/s	—	—	0.6 kW	Specs →
40	Data Center GPU Max 1100 48GB	1	61 %	516 tok/s	33.1 tok/s	—	—	0.3 kW	Specs →
41	GeForce RTX 5090 32GB	1	93 %	379 tok/s	48.3 tok/s	$1.19	$0.87	0.6 kW	Rent →
42	Ascend 910B	1	46 %	566 tok/s	43.2 tok/s	—	—	0.4 kW	Specs →
43	MI250X 128GB	1	23 %	693 tok/s	88.4 tok/s	—	—	0.6 kW	Specs →
44	RTX 6000 Ada 48GB	1	61 %	~425 tok/s	25.9 tok/s	$0.97	$0.63	0.3 kW	Rent →
45	Instinct MI250 128GB	1	23 %	664 tok/s	88.4 tok/s	—	—	0.5 kW	Specs →
46	L40S 48GB	1	61 %	401 tok/s	23.3 tok/s	$0.76	$0.53	0.3 kW	Rent →
47	A800 PCIe 80GB	1	36 %	517 tok/s	52.2 tok/s	—	—	0.3 kW	Specs →
48	A800 SXM4	1	36 %	525 tok/s	55.0 tok/s	—	—	0.4 kW	Specs →
49	A100 PCIe 80GB	1	36 %	517 tok/s	52.2 tok/s	$1.35	$0.73	0.3 kW	Rent →
50	A100 SXM4 80GB	1	36 %	525 tok/s	55.0 tok/s	$1.40	$0.74	0.4 kW	Rent →
51	A100 PCIe 40GB	1	74 %	480 tok/s	41.9 tok/s	—	—	0.3 kW	Specs →
52	A100 SXM4 40GB	1	74 %	480 tok/s	41.9 tok/s	$1.56	$0.90	0.4 kW	Rent →
53	GeForce RTX 5090 D V2 24GB	2	63 %	~838 tok/s	68.9 tok/s	—	—	1.1 kW	Specs →
54	RTX 5880 Ada 48GB	1	61 %	~369 tok/s	25.9 tok/s	—	—	0.3 kW	Specs →
55	TPU v4 32GB	1	93 %	403 tok/s	32.4 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB	1	93 %	~277 tok/s	15.5 tok/s	$0.83	$0.83	0.3 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	1	61 %	~403 tok/s	36.2 tok/s	$0.96	$0.66	0.3 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	1	41 %	~403 tok/s	36.2 tok/s	—	—	0.3 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition	1	93 %	285 tok/s	21.6 tok/s	—	—	0.2 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB	1	93 %	299 tok/s	24.2 tok/s	$0.72	$0.67	0.2 kW	Rent →
61	Radeon AI PRO R9700 32GB	1	93 %	250 tok/s	17.3 tok/s	—	—	0.3 kW	Specs →
62	Inferentia2	1	93 %	277 tok/s	22.1 tok/s	—	—	—	Specs →
63	Trainium	1	93 %	277 tok/s	22.1 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB	1	93 %	183 tok/s	33.1 tok/s	—	—	0.3 kW	Specs →
65	Arc Pro B70 32GB	1	93 %	~239 tok/s	16.4 tok/s	—	—	0.2 kW	Specs →
66	L40 48GB	1	61 %	274 tok/s	23.3 tok/s	$0.82	$0.83	0.3 kW	Rent →
67	Instinct MI210 PCIe	1	46 %	332 tok/s	44.2 tok/s	—	—	0.3 kW	Specs →
68	GeForce RTX 4090 24GB	2	63 %	517 tok/s	51.7 tok/s	$1.78	$0.96	0.9 kW	Rent →
69	A30 24GB	2	63 %	504 tok/s	47.8 tok/s	—	—	0.3 kW	Specs →
70	RTX 4500 Ada 24GB	2	63 %	~358 tok/s	22.1 tok/s	—	—	0.4 kW	Specs →
71	RTX A6000 48GB	1	61 %	238 tok/s	20.7 tok/s	$0.50	$0.58	0.3 kW	Rent →
72	A40 48GB	1	61 %	225 tok/s	18.8 tok/s	$0.59	$0.73	0.3 kW	Rent →
73	H20 141GB HBM3e	1	21 %	322 tok/s	108 tok/s	—	—	0.4 kW	Specs →
74	H20 96GB	1	30 %	322 tok/s	108 tok/s	—	—	0.4 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB	2	63 %	~417 tok/s	34.4 tok/s	$1.14	$0.76	0.3 kW	Rent →
76	BI-V100 Tiangai 100	1	93 %	263 tok/s	32.4 tok/s	—	—	0.3 kW	Specs →
77	Atlas 300I Duo 48GB	1	61 %	172 tok/s	11.0 tok/s	—	—	0.1 kW	Specs →
78	Atlas 300I Duo 96GB	1	30 %	172 tok/s	11.0 tok/s	—	—	0.1 kW	Specs →
79	RTX A5500 24GB	2	63 %	~416 tok/s	39.4 tok/s	—	—	0.5 kW	Specs →
80	Wormhole n300s	2	63 %	366 tok/s	29.5 tok/s	—	—	0.6 kW	Specs →
81	V100S PCIe 32GB	1	93 %	236 tok/s	30.6 tok/s	$0.88	$1.03	0.3 kW	Rent →
82	ATOM-Max	1	46 %	228 tok/s	27.6 tok/s	—	—	0.3 kW	Specs →
83	A10 24GB	2	63 %	361 tok/s	30.7 tok/s	$2.20	$1.70	0.3 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	1	61 %	211 tok/s	23.3 tok/s	—	—	0.3 kW	Specs →
85	Radeon PRO W7900 48GB	1	61 %	211 tok/s	23.3 tok/s	—	—	0.3 kW	Specs →
86	Radeon RX 7900 XTX 24GB	2	63 %	413 tok/s	49.2 tok/s	—	—	0.7 kW	Specs →
87	L4 24GB	2	63 %	260 tok/s	15.4 tok/s	$1.18	$1.26	0.1 kW	Rent →
88	L20 48GB	1	61 %	207 tok/s	23.3 tok/s	—	—	0.3 kW	Specs →
89	Radeon AI PRO R9600D 32GB	1	93 %	166 tok/s	17.3 tok/s	—	—	0.1 kW	Specs →
90	Arc Pro B65 32GB	1	93 %	~163 tok/s	16.4 tok/s	—	—	0.2 kW	Specs →
91	Arc Pro B60 24GB	2	63 %	~280 tok/s	23.4 tok/s	—	—	0.4 kW	Specs →
92	L2 24GB	2	63 %	232 tok/s	15.4 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB	2	63 %	~271 tok/s	22.1 tok/s	—	—	0.1 kW	Specs →
94	MLU370-X8	1	61 %	160 tok/s	16.6 tok/s	—	—	0.3 kW	Specs →
95	Radeon PRO W7800 48GB	1	61 %	168 tok/s	23.3 tok/s	—	—	0.3 kW	Specs →
96	Radeon PRO W7800 32GB	1	93 %	151 tok/s	15.5 tok/s	—	—	0.3 kW	Specs →
97	GeForce RTX 3090 24GB	2	63 %	267 tok/s	48.0 tok/s	$1.00	$1.04	0.7 kW	Rent →
98	A10G 24GB	2	63 %	241 tok/s	30.7 tok/s	$2.01	$2.32	0.6 kW	Rent →
99	Atlas 300V Pro	1	61 %	86.1 tok/s	5.5 tok/s	—	—	—	Specs →
100	Atlas 300I Pro	2	63 %	164 tok/s	10.5 tok/s	—	—	—	Specs →
101	Atlas 300V	2	63 %	136 tok/s	10.5 tok/s	—	—	0.1 kW	Specs →
102	GeForce RTX 3090 Ti 24GB	2	63 %	164 tok/s	51.7 tok/s	$0.92	$1.56	0.9 kW	Rent →
103	Radeon PRO W6800 32GB	1	93 %	71.7 tok/s	13.8 tok/s	—	—	0.3 kW	Specs →
104	RTX A5000 24GB	2	63 %	115 tok/s	39.4 tok/s	$0.54	$1.30	0.5 kW	Rent →
105	Radeon PRO V710 28GB	2	54 %	107 tok/s	23.0 tok/s	—	—	0.3 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 626 tokens/s for $0.50/hr; the AMD pick is 1 x Instinct MI300X 192GB at about 1.9k tokens/s for $2.39/hr. Per rental dollar NVIDIA delivers 1.61x the tokens of AMD here.

NVIDIA 51 parts fit

Best value: 1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 626 tok/s for $0.50/hr

Fewest GPUs: 1 x Rubin SXM, 6.4k tok/s

AMD 18 parts fit

Best value: 1 x Instinct MI300X 192GB, 1.9k tok/s for $2.39/hr

Fewest GPUs: 1 x Instinct MI455X, 7.5k tok/s

Huawei 10 parts fit

Fewest GPUs: 1 x Ascend 910C, 1.1k tok/s

Intel 7 parts fit

Best value: 1 x Gaudi 2 96GB, 695 tok/s for $1.10/hr

Fewest GPUs: 1 x Gaudi 3 128GB, 1.8k tok/s

Google 4 parts fit

Fewest GPUs: 1 x TPU v7 192GB, 3.0k tok/s

AWS 2 parts fit

Fewest GPUs: 1 x Inferentia2, 277 tok/s

Other 7 parts fit

Fewest GPUs: 1 x BR100, 884 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	1,000 x TP1	0	1.51M tok/s	$2,480	700 kW	$0.46
B	100 x GB200 NVL72 GPU 186GB	100 x TP1	0	321k tok/s	$1,050	—	$0.91

Size Nemotron 3 Nano 4B 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 Nemotron 3 Nano 4B need?

The weights take 7.9 GB at the native BF16 precision (7.9 GB at BF16, 4.0 GB at FP8, 2.2 GB at INT4). Each concurrent stream adds 16 KB of KV cache per token: 0.1 GB at 4,096 tokens and 0.5 GB at 32,768 tokens.

What is the bare minimum to run Nemotron 3 Nano 4B?

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 Nemotron 3 Nano 4B?

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 one H100, and 1,000 tokens per second takes 1 H100s across 1 replicas.

How much does Nemotron 3 Nano 4B cost per million tokens on an H100?

About $0.39 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 1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $0.22.

What precision does Nemotron 3 Nano 4B 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. Hybrid Mamba-Transformer: 4 of 42 layers keep a per-token cache (16 KB per token) and 21 Mamba-2 layers hold 84 MB of fixed FP32 state per sequence, so long contexts cost little memory. NVIDIA also publishes official FP8 and GGUF checkpoints.