NVIDIA Nemotron Nano 2

Nemotron Nano 9B v2 hardware requirements

Fits on one Instinct MI355X OAM at BF16. For 1,000 tokens/s at 32k context you need 2 x H100 SXM5 80GB at $3.66/hr or 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $2.00/hr.

Parameters

8.9B

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 Nano 9B v2

The strongest part each vendor offers for Nemotron Nano 9B v2 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 381 users, 1 x Instinct MI355X OAM 381 users, 1 x Ascend 910C 154 users, 1 x Gaudi 2 96GB 108 users, 1 x TPU v7 192GB 244 users, 1 x Inferentia2 17 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 329
Per hour: $5.99
Per million tokens: $5.06
Memory used: 6%

Rent on Daytona

Cheapest per hour

1 x RTX A5000 24GB

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

Tokens/s: 28.6
Per hour: $0.27
Per million tokens: $2.62
Memory used: 81%

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: 328
Per hour: $0.50
Per million tokens: $0.42
Memory used: 43%

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	9 %	3.9k tok/s	442 tok/s	—	—	—	Specs →
2	Rubin SXM	1	14 %	3.2k tok/s	417 tok/s	—	—	—	Specs →
3	TPU 8t announced	1	19 %	~1.8k tok/s	124 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	14 %	~1.8k tok/s	163 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	14 %	1.7k tok/s	152 tok/s	$5.99	$0.96	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	14 %	1.7k tok/s	152 tok/s	$12.42	$2.00	—	Rent →
7	GB200 NVL72 GPU 186GB	1	22 %	1.7k tok/s	152 tok/s	$10.50	$1.69	—	Rent →
8	Instinct MI350X OAM	1	14 %	1.6k tok/s	152 tok/s	$6.16	$1.04	1.0 kW	Rent →
9	TPU v7 192GB	1	21 %	1.6k tok/s	140 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	9 %	~2.1k tok/s	442 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	16 %	1.6k tok/s	152 tok/s	$4.99	$0.86	1.4 kW	Rent →
12	B200 SXM 180GB	1	23 %	1.6k tok/s	146 tok/s	$4.09	$0.71	1.0 kW	Rent →
13	Ascend 970 announced	1	14 %	1.7k tok/s	273 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	1	21 %	1.4k tok/s	152 tok/s	—	—	0.7 kW	Specs →
15	Jalapeño announced	1	19 %	~1.5k tok/s	292 tok/s	—	—	0.7 kW	Specs →
16	Instinct MI325X OAM	1	16 %	1.0k tok/s	114 tok/s	$3.80	$1.04	1.0 kW	Rent →
17	Instinct MI300X 192GB	1	21 %	976 tok/s	101 tok/s	$2.39	$0.68	0.8 kW	Rent →
18	BR100	1	65 %	515 tok/s	30.3 tok/s	—	—	0.6 kW	Specs →
19	Ascend 960 announced	1	14 %	905 tok/s	182 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	1	29 %	778 tok/s	91.0 tok/s	$2.48	$0.89	0.7 kW	Rent →
21	H100 SXM5 80GB	1	52 %	697 tok/s	63.5 tok/s	$1.83	$0.73	0.7 kW	Rent →
22	TPU v6e 32GB	2	66 %	939 tok/s	59.0 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	1	32 %	619 tok/s	62.1 tok/s	—	—	0.6 kW	Specs →
24	H200 NVL 141GB	1	29 %	~685 tok/s	91.0 tok/s	$3.61	$1.47	0.6 kW	Rent →
25	H100 NVL 94GB	1	44 %	~650 tok/s	74.0 tok/s	$3.19	$1.36	0.4 kW	Rent →
26	Ascend 910C	1	32 %	595 tok/s	60.7 tok/s	—	—	0.6 kW	Specs →
27	H100 PCIe 80GB	1	52 %	486 tok/s	37.9 tok/s	$2.50	$1.43	0.3 kW	Rent →
28	H800 PCIe 80GB	1	52 %	486 tok/s	37.9 tok/s	—	—	0.3 kW	Specs →
29	Ascend 950DT	1	29 %	472 tok/s	75.9 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	1	43 %	361 tok/s	34.0 tok/s	$1.80	$1.39	0.6 kW	Rent →
31	Ascend 950PR	1	32 %	345 tok/s	30.3 tok/s	—	—	0.9 kW	Specs →
32	MTT S5000	1	52 %	~345 tok/s	30.3 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	1	43 %	345 tok/s	30.3 tok/s	$0.69	$0.56	0.6 kW	Rent →
34	Gaudi 3 128GB	1	32 %	998 tok/s	70.2 tok/s	—	—	0.9 kW	Specs →
35	TPU v5p 95GB	1	44 %	380 tok/s	52.4 tok/s	—	—	0.5 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	1	43 %	328 tok/s	34.0 tok/s	$0.50	$0.42	0.3 kW	Rent →
37	Cloud AI 100 Ultra	1	32 %	~194 tok/s	10.4 tok/s	—	—	0.1 kW	Specs →
38	Gaudi 2 96GB	1	43 %	353 tok/s	46.5 tok/s	$1.10	$0.87	0.6 kW	Rent →
39	Atlas 350	1	37 %	296 tok/s	26.5 tok/s	—	—	0.6 kW	Specs →
40	Data Center GPU Max 1100 48GB	1	87 %	280 tok/s	23.3 tok/s	—	—	0.3 kW	Specs →
41	GeForce RTX 5090 32GB	2	66 %	354 tok/s	64.6 tok/s	$2.38	$1.87	1.1 kW	Rent →
42	Ascend 910B	1	65 %	297 tok/s	30.3 tok/s	—	—	0.4 kW	Specs →
43	MI250X 128GB	1	32 %	340 tok/s	62.1 tok/s	—	—	0.6 kW	Specs →
44	RTX 6000 Ada 48GB	1	87 %	~234 tok/s	18.2 tok/s	$0.97	$1.15	0.3 kW	Rent →
45	Instinct MI250 128GB	1	32 %	325 tok/s	62.1 tok/s	—	—	0.5 kW	Specs →
46	L40S 48GB	1	87 %	223 tok/s	16.4 tok/s	$0.76	$0.95	0.3 kW	Rent →
47	A800 PCIe 80GB	1	52 %	260 tok/s	36.7 tok/s	—	—	0.3 kW	Specs →
48	A800 SXM4	1	52 %	263 tok/s	38.7 tok/s	—	—	0.4 kW	Specs →
49	A100 PCIe 80GB	1	52 %	260 tok/s	36.7 tok/s	$1.35	$1.44	0.3 kW	Rent →
50	A100 SXM4 80GB	1	52 %	263 tok/s	38.7 tok/s	$1.40	$1.48	0.4 kW	Rent →
51	A100 PCIe 40GB	2	53 %	469 tok/s	56.0 tok/s	—	—	0.5 kW	Specs →
52	A100 SXM4 40GB	2	53 %	469 tok/s	56.0 tok/s	$3.12	$1.85	0.8 kW	Rent →
53	GeForce RTX 5090 D V2 24GB	2	89 %	~435 tok/s	48.4 tok/s	—	—	1.1 kW	Specs →
54	RTX 5880 Ada 48GB	1	87 %	~197 tok/s	18.2 tok/s	—	—	0.3 kW	Specs →
55	TPU v4 32GB	2	66 %	399 tok/s	43.2 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB	2	66 %	~295 tok/s	20.8 tok/s	$1.66	$1.56	0.5 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	1	87 %	~206 tok/s	25.5 tok/s	$0.96	$1.29	0.3 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	1	58 %	~206 tok/s	25.5 tok/s	—	—	0.3 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition	2	66 %	285 tok/s	28.8 tok/s	—	—	0.3 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB	2	66 %	295 tok/s	32.3 tok/s	$1.44	$1.36	0.4 kW	Rent →
61	Radeon AI PRO R9700 32GB	2	66 %	255 tok/s	23.1 tok/s	—	—	0.6 kW	Specs →
62	Inferentia2	2	66 %	275 tok/s	29.5 tok/s	—	—	—	Specs →
63	Trainium	2	66 %	275 tok/s	29.5 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB	2	66 %	166 tok/s	44.3 tok/s	—	—	0.6 kW	Specs →
65	Arc Pro B70 32GB	2	66 %	~244 tok/s	21.9 tok/s	—	—	0.5 kW	Specs →
66	L40 48GB	1	87 %	142 tok/s	16.4 tok/s	$0.82	$1.61	0.3 kW	Rent →
67	Instinct MI210 PCIe	1	65 %	162 tok/s	31.1 tok/s	—	—	0.3 kW	Specs →
68	GeForce RTX 4090 24GB	2	89 %	261 tok/s	36.3 tok/s	$1.78	$1.90	0.9 kW	Rent →
69	A30 24GB	2	89 %	256 tok/s	33.6 tok/s	—	—	0.3 kW	Specs →
70	RTX 4500 Ada 24GB	2	89 %	~196 tok/s	15.6 tok/s	—	—	0.4 kW	Specs →
71	RTX A6000 48GB	1	87 %	122 tok/s	14.6 tok/s	$0.50	$1.14	0.3 kW	Rent →
72	A40 48GB	1	87 %	116 tok/s	13.2 tok/s	$0.59	$1.41	0.3 kW	Rent →
73	H20 141GB HBM3e	1	29 %	149 tok/s	75.9 tok/s	—	—	0.4 kW	Specs →
74	H20 96GB	1	43 %	149 tok/s	75.9 tok/s	—	—	0.4 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB	2	89 %	~216 tok/s	24.2 tok/s	$1.14	$1.47	0.3 kW	Rent →
76	BI-V100 Tiangai 100	2	66 %	246 tok/s	43.2 tok/s	—	—	0.5 kW	Specs →
77	Atlas 300I Duo 48GB	1	87 %	93.4 tok/s	7.7 tok/s	—	—	0.1 kW	Specs →
78	Atlas 300I Duo 96GB	1	43 %	93.4 tok/s	7.7 tok/s	—	—	0.1 kW	Specs →
79	RTX A5500 24GB	2	89 %	~211 tok/s	27.7 tok/s	—	—	0.5 kW	Specs →
80	Wormhole n300s	2	89 %	190 tok/s	20.8 tok/s	—	—	0.6 kW	Specs →
81	V100S PCIe 32GB	2	66 %	220 tok/s	40.9 tok/s	$1.76	$2.22	0.5 kW	Rent →
82	ATOM-Max	1	65 %	112 tok/s	19.4 tok/s	—	—	0.3 kW	Specs →
83	A10 24GB	2	89 %	186 tok/s	21.6 tok/s	$2.20	$3.29	0.3 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	1	87 %	105 tok/s	16.4 tok/s	—	—	0.3 kW	Specs →
85	Radeon PRO W7900 48GB	1	87 %	105 tok/s	16.4 tok/s	—	—	0.3 kW	Specs →
86	Radeon RX 7900 XTX 24GB	2	89 %	204 tok/s	34.6 tok/s	—	—	0.7 kW	Specs →
87	L4 24GB	2	89 %	144 tok/s	10.8 tok/s	$1.18	$2.28	0.1 kW	Rent →
88	L20 48GB	1	87 %	103 tok/s	16.4 tok/s	—	—	0.3 kW	Specs →
89	Radeon AI PRO R9600D 32GB	2	66 %	158 tok/s	23.1 tok/s	—	—	0.3 kW	Specs →
90	Arc Pro B65 32GB	2	66 %	~156 tok/s	21.9 tok/s	—	—	0.4 kW	Specs →
91	Arc Pro B60 24GB	2	89 %	~145 tok/s	16.4 tok/s	—	—	0.4 kW	Specs →
92	L2 24GB	2	89 %	125 tok/s	10.8 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB	2	89 %	~141 tok/s	15.6 tok/s	—	—	0.1 kW	Specs →
94	MLU370-X8	1	87 %	80.6 tok/s	11.7 tok/s	—	—	0.3 kW	Specs →
95	Radeon PRO W7800 48GB	1	87 %	81.8 tok/s	16.4 tok/s	—	—	0.3 kW	Specs →
96	Radeon PRO W7800 32GB	2	66 %	144 tok/s	20.8 tok/s	—	—	0.5 kW	Specs →
97	GeForce RTX 3090 24GB	2	89 %	127 tok/s	33.7 tok/s	$1.00	$2.18	0.7 kW	Rent →
98	A10G 24GB	2	89 %	118 tok/s	21.6 tok/s	$2.01	$4.73	0.6 kW	Rent →
99	Atlas 300V Pro	1	87 %	46.8 tok/s	3.9 tok/s	—	—	—	Specs →
100	Atlas 300I Pro	2	89 %	88.9 tok/s	7.4 tok/s	—	—	—	Specs →
101	Atlas 300V	2	89 %	71.1 tok/s	7.4 tok/s	—	—	0.1 kW	Specs →
102	GeForce RTX 3090 Ti 24GB	2	89 %	76.1 tok/s	36.3 tok/s	$0.92	$3.36	0.9 kW	Rent →
103	Radeon PRO W6800 32GB	2	66 %	64.7 tok/s	18.4 tok/s	—	—	0.5 kW	Specs →
104	RTX A5000 24GB	2	89 %	53.2 tok/s	27.7 tok/s	$0.54	$2.82	0.5 kW	Rent →
105	Radeon PRO V710 28GB	2	76 %	50.8 tok/s	16.1 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 328 tokens/s for $0.50/hr; the AMD pick is 1 x Instinct MI300X 192GB at about 976 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, 328 tok/s for $0.50/hr

Fewest GPUs: 1 x Rubin SXM, 3.2k tok/s

AMD 18 parts fit

Best value: 1 x Instinct MI300X 192GB, 976 tok/s for $2.39/hr

Fewest GPUs: 1 x Instinct MI455X, 3.9k tok/s

Huawei 10 parts fit

Fewest GPUs: 1 x Ascend 910C, 595 tok/s

Intel 7 parts fit

Best value: 1 x Gaudi 2 96GB, 353 tok/s for $1.10/hr

Fewest GPUs: 1 x Gaudi 3 128GB, 998 tok/s

Google 4 parts fit

Fewest GPUs: 1 x TPU v7 192GB, 1.6k tok/s

AWS 2 parts fit

Fewest GPUs: 2 x Inferentia2, 275 tok/s

Other 7 parts fit

Fewest GPUs: 1 x BR100, 515 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	778k tok/s	$2,480	700 kW	$0.89
B	100 x GB200 NVL72 GPU 186GB	100 x TP1	0	173k tok/s	$1,050	—	$1.69

Size Nemotron Nano 9B v2 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 Nano 9B v2 need?

The weights take 17.8 GB at the native BF16 precision (17.8 GB at BF16, 8.9 GB at FP8, 5.0 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 Nano 9B v2?

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 Nano 9B v2?

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 2 H100s across 2 replicas.

How much does Nemotron Nano 9B v2 cost per million tokens on an H100?

About $0.73 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 $0.42.

What precision does Nemotron Nano 9B v2 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 56 layers keep a per-token cache (16 KB per token) and 27 Mamba-2 layers hold 144 MB of fixed state per sequence. NVIDIA's card asks servers to keep that state in FP32 for accuracy.