NVIDIA Nemotron 3

Nemotron 3 Super 120B-A12B hardware requirements

Fits on one Instinct MI355X OAM at BF16. For 1,000 tokens/s at 32k context you need 8 x H100 SXM5 80GB at $14.64/hr or 12 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $6.00/hr.

Parameters

124B

Active per token

12.0B

mixture of experts

KV cache per token

8.0 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Nemotron 3 Super 120B-A12B

The strongest part each vendor offers for Nemotron 3 Super 120B-A12B 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.

2 x GB200 NVL72 GPU 186GB 1 x Instinct MI355X OAM 4 x Ascend 910C 4 x Gaudi 2 96GB 2 x TPU v7 192GB 16 x Inferentia2

Memory fills at: 2 x GB200 NVL72 GPU 186GB 251 users, 1 x Instinct MI355X OAM 68 users, 4 x Ascend 910C 554 users, 4 x Gaudi 2 96GB 272 users, 2 x TPU v7 192GB 277 users, 16 x Inferentia2 525 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 245
Per hour: $5.99
Per million tokens: $6.80
Memory used: 89%

Rent on Daytona

Cheapest per hour

4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

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

Tokens/s: 196
Per hour: $2.00
Per million tokens: $2.83
Memory used: 67%

Rent on RunPod

Cheapest per million tokens

4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 473
Per hour: $2.00
Per million tokens: $1.18
Memory used: 71%

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	63 %	1.6k tok/s	85.6 tok/s	—	—	—	Specs →
2	Rubin SXM	1	94 %	1.4k tok/s	80.9 tok/s	—	—	—	Specs →
3	TPU 8t announced	2	63 %	~1.1k tok/s	45.6 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	94 %	~672 tok/s	31.6 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	94 %	640 tok/s	29.4 tok/s	$5.99	$2.60	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	94 %	638 tok/s	29.4 tok/s	$12.42	$5.41	—	Rent →
7	GB200 NVL72 GPU 186GB	2	73 %	1.2k tok/s	55.9 tok/s	$21.00	$4.81	—	Rent →
8	Instinct MI350X OAM	1	94 %	622 tok/s	29.4 tok/s	$6.16	$2.75	1.0 kW	Rent →
9	TPU v7 192GB	2	71 %	1.1k tok/s	51.5 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	63 %	~1.1k tok/s	85.6 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	2	52 %	1.2k tok/s	55.9 tok/s	$9.98	$2.37	2.8 kW	Rent →
12	B200 SXM 180GB	2	76 %	1.1k tok/s	53.8 tok/s	$8.18	$1.99	2.0 kW	Rent →
13	Ascend 970 announced	1	94 %	819 tok/s	52.9 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	2	71 %	1.1k tok/s	55.9 tok/s	—	—	1.4 kW	Specs →
15	Jalapeño announced	2	63 %	~1.5k tok/s	108 tok/s	—	—	1.4 kW	Specs →
16	Instinct MI325X OAM	2	53 %	798 tok/s	41.9 tok/s	$7.60	$2.65	2.0 kW	Rent →
17	Instinct MI300X 192GB	2	71 %	740 tok/s	37.0 tok/s	$4.78	$1.79	1.5 kW	Rent →
18	BR100	8	54 %	1.0k tok/s	40.0 tok/s	—	—	4.4 kW	Specs →
19	Ascend 960 announced	1	94 %	466 tok/s	35.3 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	2	97 %	624 tok/s	33.5 tok/s	$4.96	$2.21	1.4 kW	Rent →
21	H100 SXM5 80GB	4	86 %	945 tok/s	44.3 tok/s	$7.32	$2.15	2.8 kW	Rent →
22	TPU v6e 32GB 2 nodes	16	55 %	1.8k tok/s	72.2 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	4	53 %	879 tok/s	43.4 tok/s	—	—	2.4 kW	Specs →
24	H200 NVL 141GB	2	97 %	~579 tok/s	33.5 tok/s	$7.23	$3.47	1.2 kW	Rent →
25	H100 NVL 94GB	4	73 %	~976 tok/s	51.6 tok/s	$12.76	$3.63	1.6 kW	Rent →
26	Ascend 910C	4	53 %	851 tok/s	42.3 tok/s	—	—	2.4 kW	Specs →
27	H100 PCIe 80GB	4	86 %	608 tok/s	26.5 tok/s	$10.00	$4.57	1.4 kW	Rent →
28	H800 PCIe 80GB	4	86 %	608 tok/s	26.5 tok/s	—	—	1.4 kW	Specs →
29	Ascend 950DT	2	95 %	429 tok/s	27.9 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	4	71 %	497 tok/s	23.7 tok/s	$7.20	$4.03	2.4 kW	Rent →
31	Ascend 950PR	4	53 %	460 tok/s	21.2 tok/s	—	—	3.6 kW	Specs →
32	MTT S5000	4	86 %	~460 tok/s	21.2 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	4	71 %	459 tok/s	21.1 tok/s	$2.76	$1.67	2.4 kW	Rent →
34	Gaudi 3 128GB	4	53 %	1.2k tok/s	49.0 tok/s	—	—	3.6 kW	Specs →
35	TPU v5p 95GB	4	72 %	619 tok/s	36.6 tok/s	—	—	1.8 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	4	71 %	473 tok/s	23.7 tok/s	$2.00	$1.18	1.2 kW	Rent →
37	Cloud AI 100 Ultra	4	53 %	~196 tok/s	7.3 tok/s	—	—	0.6 kW	Specs →
38	Gaudi 2 96GB	4	71 %	564 tok/s	32.4 tok/s	$4.40	$2.17	2.4 kW	Rent →
39	Atlas 350	4	61 %	398 tok/s	18.5 tok/s	—	—	2.4 kW	Specs →
40	Data Center GPU Max 1100 48GB	8	72 %	685 tok/s	30.7 tok/s	—	—	2.4 kW	Specs →
41	GeForce RTX 5090 32GB 2 nodes	16	55 %	1.1k tok/s	79.0 tok/s	$19.04	$4.74	9.2 kW	Rent →
42	Ascend 910B	8	54 %	803 tok/s	40.0 tok/s	—	—	3.2 kW	Specs →
43	MI250X 128GB	4	53 %	612 tok/s	43.4 tok/s	—	—	2.2 kW	Specs →
44	RTX 6000 Ada 48GB	8	72 %	~552 tok/s	24.0 tok/s	$7.76	$3.90	2.4 kW	Rent →
45	Instinct MI250 128GB	4	53 %	593 tok/s	43.4 tok/s	—	—	2.0 kW	Specs →
46	L40S 48GB	8	72 %	510 tok/s	21.6 tok/s	$6.08	$3.31	2.8 kW	Rent →
47	A800 PCIe 80GB	4	86 %	427 tok/s	25.6 tok/s	—	—	1.2 kW	Specs →
48	A800 SXM4	4	86 %	439 tok/s	27.0 tok/s	—	—	1.6 kW	Specs →
49	A100 PCIe 80GB	4	86 %	427 tok/s	25.6 tok/s	$5.40	$3.51	1.2 kW	Rent →
50	A100 SXM4 80GB	4	86 %	439 tok/s	27.0 tok/s	$5.60	$3.55	1.6 kW	Rent →
51	A100 PCIe 40GB	8	87 %	715 tok/s	38.9 tok/s	—	—	2.0 kW	Specs →
52	A100 SXM4 40GB	8	87 %	715 tok/s	38.9 tok/s	$12.49	$4.85	3.2 kW	Rent →
53	GeForce RTX 5090 D V2 24GB 2 nodes	16	74 %	~1.1k tok/s	59.3 tok/s	—	—	9.2 kW	Specs →
54	RTX 5880 Ada 48GB	8	72 %	~507 tok/s	24.0 tok/s	—	—	2.3 kW	Specs →
55	TPU v4 32GB 2 nodes	16	55 %	1.0k tok/s	52.9 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB 2 nodes	16	55 %	~613 tok/s	25.4 tok/s	$13.28	$6.02	4.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	8	72 %	~607 tok/s	33.6 tok/s	$7.68	$3.52	2.4 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	4	95 %	~321 tok/s	17.8 tok/s	—	—	1.2 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition 2 nodes	16	55 %	713 tok/s	35.3 tok/s	—	—	2.6 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB 2 nodes	16	55 %	765 tok/s	39.5 tok/s	$11.52	$4.19	3.2 kW	Rent →
61	Radeon AI PRO R9700 32GB 2 nodes	16	55 %	604 tok/s	28.2 tok/s	—	—	4.8 kW	Specs →
62	Inferentia2 2 nodes	16	55 %	706 tok/s	36.2 tok/s	—	—	—	Specs →
63	Trainium 2 nodes	16	55 %	706 tok/s	36.2 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB 2 nodes	16	55 %	583 tok/s	54.2 tok/s	—	—	4.8 kW	Specs →
65	Arc Pro B70 32GB 2 nodes	16	55 %	~576 tok/s	26.8 tok/s	—	—	3.7 kW	Specs →
66	L40 48GB	8	72 %	405 tok/s	21.6 tok/s	$6.56	$4.50	2.4 kW	Rent →
67	Instinct MI210 PCIe	8	54 %	560 tok/s	40.9 tok/s	—	—	2.4 kW	Specs →
68	GeForce RTX 4090 24GB 2 nodes	16	74 %	747 tok/s	44.5 tok/s	$14.24	$5.30	7.2 kW	Rent →
69	A30 24GB 2 nodes	16	74 %	716 tok/s	41.2 tok/s	—	—	2.6 kW	Specs →
70	RTX 4500 Ada 24GB 2 nodes	16	74 %	~434 tok/s	19.1 tok/s	—	—	3.4 kW	Specs →
71	RTX A6000 48GB	8	72 %	354 tok/s	19.2 tok/s	$4.00	$3.14	2.4 kW	Rent →
72	A40 48GB	8	72 %	330 tok/s	17.4 tok/s	$4.72	$3.98	2.4 kW	Rent →
73	H20 141GB HBM3e	2	97 %	179 tok/s	27.9 tok/s	—	—	0.8 kW	Specs →
74	H20 96GB	4	71 %	338 tok/s	52.9 tok/s	—	—	1.6 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB 2 nodes	16	74 %	~566 tok/s	29.6 tok/s	$9.12	$4.48	2.3 kW	Rent →
76	BI-V100 Tiangai 100 2 nodes	16	55 %	766 tok/s	52.9 tok/s	—	—	4.0 kW	Specs →
77	Atlas 300I Duo 48GB	8	72 %	228 tok/s	10.2 tok/s	—	—	1.2 kW	Specs →
78	Atlas 300I Duo 96GB	4	71 %	121 tok/s	5.4 tok/s	—	—	0.6 kW	Specs →
79	RTX A5500 24GB 2 nodes	16	74 %	~591 tok/s	33.9 tok/s	—	—	3.7 kW	Specs →
80	Wormhole n300s 2 nodes	16	74 %	492 tok/s	25.4 tok/s	—	—	4.8 kW	Specs →
81	V100S PCIe 32GB 2 nodes	16	55 %	698 tok/s	50.0 tok/s	$14.08	$5.60	4.0 kW	Rent →
82	ATOM-Max	8	54 %	375 tok/s	25.6 tok/s	—	—	2.8 kW	Specs →
83	A10 24GB 2 nodes	16	74 %	495 tok/s	26.5 tok/s	$17.63	$9.89	2.4 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	8	72 %	339 tok/s	21.6 tok/s	—	—	2.4 kW	Specs →
85	Radeon PRO W7900 48GB	8	72 %	339 tok/s	21.6 tok/s	—	—	2.4 kW	Specs →
86	Radeon RX 7900 XTX 24GB 2 nodes	16	74 %	628 tok/s	42.3 tok/s	—	—	5.7 kW	Specs →
87	L4 24GB 2 nodes	16	74 %	310 tok/s	13.2 tok/s	$9.44	$8.47	1.2 kW	Rent →
88	L20 48GB	8	72 %	334 tok/s	21.6 tok/s	—	—	2.2 kW	Specs →
89	Radeon AI PRO R9600D 32GB 2 nodes	16	55 %	461 tok/s	28.2 tok/s	—	—	2.4 kW	Specs →
90	Arc Pro B65 32GB 2 nodes	16	55 %	~448 tok/s	26.8 tok/s	—	—	3.2 kW	Specs →
91	Arc Pro B60 24GB 2 nodes	16	74 %	~382 tok/s	20.1 tok/s	—	—	3.2 kW	Specs →
92	L2 24GB 2 nodes	16	74 %	290 tok/s	13.2 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB 2 nodes	16	74 %	~366 tok/s	19.1 tok/s	—	—	1.1 kW	Specs →
94	MLU370-X8	8	72 %	252 tok/s	15.4 tok/s	—	—	2.0 kW	Specs →
95	Radeon PRO W7800 48GB	8	72 %	286 tok/s	21.6 tok/s	—	—	2.1 kW	Specs →
96	Radeon PRO W7800 32GB 2 nodes	16	55 %	418 tok/s	25.4 tok/s	—	—	4.2 kW	Specs →
97	GeForce RTX 3090 24GB 2 nodes	16	74 %	447 tok/s	41.3 tok/s	$8.00	$4.97	5.6 kW	Rent →
98	A10G 24GB 2 nodes	16	74 %	373 tok/s	26.5 tok/s	$16.10	$11.99	4.8 kW	Rent →
99	Atlas 300V Pro	8	72 %	114 tok/s	5.1 tok/s	—	—	—	Specs →
100	Atlas 300I Pro 2 nodes	16	74 %	202 tok/s	9.0 tok/s	—	—	—	Specs →
101	Atlas 300V 2 nodes	16	74 %	180 tok/s	9.0 tok/s	—	—	1.2 kW	Specs →
102	GeForce RTX 3090 Ti 24GB 2 nodes	16	74 %	300 tok/s	44.5 tok/s	$7.36	$6.80	7.2 kW	Rent →
103	Radeon PRO W6800 32GB 2 nodes	16	55 %	231 tok/s	22.6 tok/s	—	—	4.0 kW	Specs →
104	RTX A5000 24GB 2 nodes	16	74 %	213 tok/s	33.9 tok/s	$4.32	$5.64	3.7 kW	Rent →
105	Radeon PRO V710 28GB 2 nodes	16	63 %	186 tok/s	19.8 tok/s	—	—	2.5 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 473 tokens/s for $2.00/hr; the AMD pick is 2 x Instinct MI300X 192GB at about 740 tokens/s for $4.78/hr. Per rental dollar NVIDIA delivers 1.53x the tokens of AMD here.

NVIDIA 51 parts fit

Best value: 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 473 tok/s for $2.00/hr

Fewest GPUs: 1 x Rubin SXM, 1.4k tok/s

AMD 18 parts fit

Best value: 2 x Instinct MI300X 192GB, 740 tok/s for $4.78/hr

Fewest GPUs: 1 x Instinct MI455X, 1.6k tok/s

Huawei 10 parts fit

Fewest GPUs: 2 x Ascend 950DT, 429 tok/s

Intel 7 parts fit

Best value: 4 x Gaudi 2 96GB, 564 tok/s for $4.40/hr

Fewest GPUs: 4 x Gaudi 3 128GB, 1.2k tok/s

Google 4 parts fit

Fewest GPUs: 2 x TPU v7 192GB, 1.1k tok/s

AWS 2 parts fit

Fewest GPUs: 16 x Inferentia2, 706 tok/s

Other 7 parts fit

Fewest GPUs: 4 x MTT S5000, 460 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	500 x TP2	0	312k tok/s	$2,480	700 kW	$2.21
B	100 x GB200 NVL72 GPU 186GB	50 x TP2	0	61k tok/s	$1,050	—	$4.81

Size Nemotron 3 Super 120B-A12B 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 Super 120B-A12B need?

The weights take 247 GB at the native BF16 precision (247 GB at BF16, 124 GB at FP8, 69.5 GB at INT4). Each concurrent stream adds 8.0 KB of KV cache per token: 0.0 GB at 4,096 tokens and 0.3 GB at 32,768 tokens.

What is the bare minimum to run Nemotron 3 Super 120B-A12B?

1 x Instinct MI355X OAM at BF16 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $2.00 per hour.

How many H100s do you need to run Nemotron 3 Super 120B-A12B?

4 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 4 H100s, and 1,000 tokens per second takes 8 H100s across 2 replicas.

How much does Nemotron 3 Super 120B-A12B cost per million tokens on an H100?

About $2.15 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 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.18.

What precision does Nemotron 3 Super 120B-A12B 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 with latent-width MoE experts and a multi-token-prediction layer. 8 attention layers keep a per-token cache (8 KB per token); 40 Mamba-2 layers hold 170 MB of fixed FP32 state per sequence. NVIDIA states the BF16 checkpoint fits on 2 B200 or B300 and the NVFP4 checkpoint on a single B200 or DGX Spark.