NVIDIA Nemotron 3.5

Nemotron 3.5 Lightning 30B-A3B 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 3 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.50/hr.

Parameters

32.9B

Active per token

3.0B

mixture of experts

KV cache per token

6.0 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Nemotron 3.5 Lightning 30B-A3B

The strongest part each vendor offers for Nemotron 3.5 Lightning 30B-A3B 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 4 x Inferentia2

Memory fills at: 1 x GB300 NVL72 GPU 288GB 844 users, 1 x Instinct MI355X OAM 844 users, 1 x Ascend 910C 226 users, 1 x Gaudi 2 96GB 103 users, 1 x TPU v7 192GB 473 users, 4 x Inferentia2 213 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 974
Per hour: $5.99
Per million tokens: $1.71
Memory used: 24%

Rent on Daytona

Cheapest per hour

1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

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

Tokens/s: 217
Per hour: $0.50
Per million tokens: $0.64
Memory used: 72%

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: 472
Per hour: $0.50
Per million tokens: $0.29
Memory used: 81%

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	18 %	5.9k tok/s	290 tok/s	—	—	—	Specs →
2	Rubin SXM	1	27 %	5.2k tok/s	273 tok/s	—	—	—	Specs →
3	TPU 8t announced	1	36 %	~2.1k tok/s	81.1 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	27 %	~2.4k tok/s	107 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	27 %	2.3k tok/s	99.4 tok/s	$5.99	$0.73	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	27 %	2.3k tok/s	99.4 tok/s	$12.42	$1.52	—	Rent →
7	GB200 NVL72 GPU 186GB	1	41 %	2.3k tok/s	99.4 tok/s	$10.50	$1.28	—	Rent →
8	Instinct MI350X OAM	1	27 %	2.2k tok/s	99.4 tok/s	$6.16	$0.77	1.0 kW	Rent →
9	TPU v7 192GB	1	40 %	2.1k tok/s	91.7 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	18 %	~4.1k tok/s	290 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	29 %	2.2k tok/s	99.4 tok/s	$4.99	$0.63	1.4 kW	Rent →
12	B200 SXM 180GB	1	43 %	2.1k tok/s	95.7 tok/s	$4.09	$0.53	1.0 kW	Rent →
13	Ascend 970 announced	1	27 %	3.0k tok/s	179 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	1	40 %	2.0k tok/s	99.4 tok/s	—	—	0.7 kW	Specs →
15	Jalapeño announced	1	36 %	~2.9k tok/s	191 tok/s	—	—	0.7 kW	Specs →
16	Instinct MI325X OAM	1	30 %	1.5k tok/s	74.6 tok/s	$3.80	$0.70	1.0 kW	Rent →
17	Instinct MI300X 192GB	1	40 %	1.4k tok/s	65.9 tok/s	$2.39	$0.47	0.8 kW	Rent →
18	BR100	2	61 %	1.0k tok/s	37.8 tok/s	—	—	1.1 kW	Specs →
19	Ascend 960 announced	1	27 %	1.7k tok/s	119 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	1	55 %	1.2k tok/s	59.7 tok/s	$2.48	$0.58	0.7 kW	Rent →
21	H100 SXM5 80GB	1	97 %	936 tok/s	41.6 tok/s	$1.83	$0.54	0.7 kW	Rent →
22	TPU v6e 32GB	4	62 %	1.9k tok/s	73.3 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	1	60 %	875 tok/s	40.7 tok/s	—	—	0.6 kW	Specs →
24	H200 NVL 141GB	1	55 %	~1.1k tok/s	59.7 tok/s	$3.61	$0.90	0.6 kW	Rent →
25	H100 NVL 94GB	1	82 %	~979 tok/s	48.5 tok/s	$3.19	$0.91	0.4 kW	Rent →
26	Ascend 910C	1	60 %	848 tok/s	39.8 tok/s	—	—	0.6 kW	Specs →
27	H100 PCIe 80GB	1	97 %	598 tok/s	24.9 tok/s	$2.50	$1.16	0.3 kW	Rent →
28	H800 PCIe 80GB	1	97 %	598 tok/s	24.9 tok/s	—	—	0.3 kW	Specs →
29	Ascend 950DT	1	53 %	830 tok/s	49.7 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	1	81 %	493 tok/s	22.3 tok/s	$1.80	$1.01	0.6 kW	Rent →
31	Ascend 950PR	1	60 %	454 tok/s	19.9 tok/s	—	—	0.9 kW	Specs →
32	MTT S5000	1	97 %	~454 tok/s	19.9 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	1	81 %	454 tok/s	19.9 tok/s	$0.69	$0.42	0.6 kW	Rent →
34	Gaudi 3 128GB	1	60 %	1.2k tok/s	46.0 tok/s	—	—	0.9 kW	Specs →
35	TPU v5p 95GB	1	81 %	627 tok/s	34.4 tok/s	—	—	0.5 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	1	81 %	472 tok/s	22.3 tok/s	$0.50	$0.29	0.3 kW	Rent →
37	Cloud AI 100 Ultra	1	60 %	~188 tok/s	6.8 tok/s	—	—	0.1 kW	Specs →
38	Gaudi 2 96GB	1	81 %	570 tok/s	30.5 tok/s	$1.10	$0.54	0.6 kW	Rent →
39	Atlas 350	1	69 %	394 tok/s	17.4 tok/s	—	—	0.6 kW	Specs →
40	Data Center GPU Max 1100 48GB	2	81 %	679 tok/s	29.0 tok/s	—	—	0.6 kW	Specs →
41	GeForce RTX 5090 32GB	4	62 %	1.2k tok/s	80.2 tok/s	$4.76	$1.07	2.3 kW	Rent →
42	Ascend 910B	2	61 %	806 tok/s	37.8 tok/s	—	—	0.8 kW	Specs →
43	MI250X 128GB	1	60 %	629 tok/s	40.7 tok/s	—	—	0.6 kW	Specs →
44	RTX 6000 Ada 48GB	2	81 %	~545 tok/s	22.7 tok/s	$1.94	$0.99	0.6 kW	Rent →
45	Instinct MI250 128GB	1	60 %	611 tok/s	40.7 tok/s	—	—	0.5 kW	Specs →
46	L40S 48GB	2	81 %	503 tok/s	20.4 tok/s	$1.52	$0.84	0.7 kW	Rent →
47	A800 PCIe 80GB	1	97 %	433 tok/s	24.1 tok/s	—	—	0.3 kW	Specs →
48	A800 SXM4	1	97 %	446 tok/s	25.3 tok/s	—	—	0.4 kW	Specs →
49	A100 PCIe 80GB	1	97 %	433 tok/s	24.1 tok/s	$1.35	$0.87	0.3 kW	Rent →
50	A100 SXM4 80GB	1	97 %	446 tok/s	25.3 tok/s	$1.40	$0.87	0.4 kW	Rent →
51	A100 PCIe 40GB	2	98 %	723 tok/s	36.7 tok/s	—	—	0.5 kW	Specs →
52	A100 SXM4 40GB	2	98 %	723 tok/s	36.7 tok/s	$3.12	$1.20	0.8 kW	Rent →
53	GeForce RTX 5090 D V2 24GB	4	83 %	~1.2k tok/s	60.1 tok/s	—	—	2.3 kW	Specs →
54	RTX 5880 Ada 48GB	2	81 %	~506 tok/s	22.7 tok/s	—	—	0.6 kW	Specs →
55	TPU v4 32GB	4	62 %	1.1k tok/s	53.7 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB	4	62 %	~646 tok/s	25.8 tok/s	$3.32	$1.43	1.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	2	81 %	~615 tok/s	31.7 tok/s	$1.92	$0.87	0.6 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	2	54 %	~615 tok/s	31.7 tok/s	—	—	0.6 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition	4	62 %	767 tok/s	35.8 tok/s	—	—	0.7 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB	4	62 %	826 tok/s	40.1 tok/s	$2.88	$0.97	0.8 kW	Rent →
61	Radeon AI PRO R9700 32GB	4	62 %	645 tok/s	28.6 tok/s	—	—	1.2 kW	Specs →
62	Inferentia2	4	62 %	762 tok/s	36.7 tok/s	—	—	—	Specs →
63	Trainium	4	62 %	762 tok/s	36.7 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB	4	62 %	660 tok/s	55.0 tok/s	—	—	1.2 kW	Specs →
65	Arc Pro B70 32GB	4	62 %	~615 tok/s	27.2 tok/s	—	—	0.9 kW	Specs →
66	L40 48GB	2	81 %	409 tok/s	20.4 tok/s	$1.64	$1.11	0.6 kW	Rent →
67	Instinct MI210 PCIe	2	61 %	580 tok/s	38.7 tok/s	—	—	0.6 kW	Specs →
68	GeForce RTX 4090 24GB	4	83 %	818 tok/s	45.1 tok/s	$3.56	$1.21	1.8 kW	Rent →
69	A30 24GB	4	83 %	782 tok/s	41.8 tok/s	—	—	0.7 kW	Specs →
70	RTX 4500 Ada 24GB	4	83 %	~461 tok/s	19.3 tok/s	—	—	0.8 kW	Specs →
71	RTX A6000 48GB	2	81 %	358 tok/s	18.1 tok/s	$1.00	$0.78	0.6 kW	Rent →
72	A40 48GB	2	81 %	332 tok/s	16.4 tok/s	$1.18	$0.99	0.6 kW	Rent →
73	H20 141GB HBM3e	1	55 %	363 tok/s	49.7 tok/s	—	—	0.4 kW	Specs →
74	H20 96GB	1	81 %	363 tok/s	49.7 tok/s	—	—	0.4 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB	4	83 %	~612 tok/s	30.1 tok/s	$2.28	$1.03	0.6 kW	Rent →
76	BI-V100 Tiangai 100	4	62 %	850 tok/s	53.7 tok/s	—	—	1.0 kW	Specs →
77	Atlas 300I Duo 48GB	2	81 %	226 tok/s	9.6 tok/s	—	—	0.3 kW	Specs →
78	Atlas 300I Duo 96GB	1	81 %	119 tok/s	5.1 tok/s	—	—	0.1 kW	Specs →
79	RTX A5500 24GB	4	83 %	~645 tok/s	34.4 tok/s	—	—	0.9 kW	Specs →
80	Wormhole n300s	4	83 %	532 tok/s	25.8 tok/s	—	—	1.2 kW	Specs →
81	V100S PCIe 32GB	4	62 %	776 tok/s	50.7 tok/s	$3.52	$1.26	1.0 kW	Rent →
82	ATOM-Max	2	61 %	387 tok/s	24.2 tok/s	—	—	0.7 kW	Specs →
83	A10 24GB	4	83 %	537 tok/s	26.9 tok/s	$4.41	$2.28	0.6 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	2	81 %	347 tok/s	20.4 tok/s	—	—	0.6 kW	Specs →
85	Radeon PRO W7900 48GB	2	81 %	347 tok/s	20.4 tok/s	—	—	0.6 kW	Specs →
86	Radeon RX 7900 XTX 24GB	4	83 %	694 tok/s	43.0 tok/s	—	—	1.4 kW	Specs →
87	L4 24GB	4	83 %	328 tok/s	13.4 tok/s	$2.36	$2.00	0.3 kW	Rent →
88	L20 48GB	2	81 %	343 tok/s	20.4 tok/s	—	—	0.6 kW	Specs →
89	Radeon AI PRO R9600D 32GB	4	62 %	506 tok/s	28.6 tok/s	—	—	0.6 kW	Specs →
90	Arc Pro B65 32GB	4	62 %	~491 tok/s	27.2 tok/s	—	—	0.8 kW	Specs →
91	Arc Pro B60 24GB	4	83 %	~413 tok/s	20.4 tok/s	—	—	0.8 kW	Specs →
92	L2 24GB	4	83 %	309 tok/s	13.4 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB	4	83 %	~396 tok/s	19.3 tok/s	—	—	0.3 kW	Specs →
94	MLU370-X8	2	81 %	258 tok/s	14.5 tok/s	—	—	0.5 kW	Specs →
95	Radeon PRO W7800 48GB	2	81 %	297 tok/s	20.4 tok/s	—	—	0.5 kW	Specs →
96	Radeon PRO W7800 32GB	4	62 %	459 tok/s	25.8 tok/s	—	—	1.0 kW	Specs →
97	GeForce RTX 3090 24GB	4	83 %	505 tok/s	41.9 tok/s	$2.00	$1.10	1.4 kW	Rent →
98	A10G 24GB	4	83 %	414 tok/s	26.9 tok/s	$4.02	$2.70	1.2 kW	Rent →
99	Atlas 300V Pro	2	81 %	113 tok/s	4.8 tok/s	—	—	—	Specs →
100	Atlas 300I Pro	4	83 %	215 tok/s	9.2 tok/s	—	—	—	Specs →
101	Atlas 300V	4	83 %	194 tok/s	9.2 tok/s	—	—	0.3 kW	Specs →
102	GeForce RTX 3090 Ti 24GB	4	83 %	347 tok/s	45.1 tok/s	$1.84	$1.47	1.8 kW	Rent →
103	Radeon PRO W6800 32GB	4	62 %	262 tok/s	22.9 tok/s	—	—	1.0 kW	Specs →
104	RTX A5000 24GB	4	83 %	247 tok/s	34.4 tok/s	$1.08	$1.22	0.9 kW	Rent →
105	Radeon PRO V710 28GB	4	71 %	212 tok/s	20.0 tok/s	—	—	0.6 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 472 tokens/s for $0.50/hr; the AMD pick is 1 x Instinct MI300X 192GB at about 1.4k 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, 472 tok/s for $0.50/hr

Fewest GPUs: 1 x Rubin SXM, 5.2k tok/s

AMD 18 parts fit

Best value: 1 x Instinct MI300X 192GB, 1.4k tok/s for $2.39/hr

Fewest GPUs: 1 x Instinct MI455X, 5.9k tok/s

Huawei 10 parts fit

Fewest GPUs: 1 x Ascend 910C, 848 tok/s

Intel 7 parts fit

Best value: 1 x Gaudi 2 96GB, 570 tok/s for $1.10/hr

Fewest GPUs: 1 x Gaudi 3 128GB, 1.2k tok/s

Google 4 parts fit

Fewest GPUs: 1 x TPU v7 192GB, 2.1k tok/s

AWS 2 parts fit

Fewest GPUs: 4 x Inferentia2, 762 tok/s

Other 7 parts fit

Fewest GPUs: 1 x MTT S5000, 454 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	1.19M tok/s	$2,480	700 kW	$0.58
B	100 x GB200 NVL72 GPU 186GB	100 x TP1	0	227k tok/s	$1,050	—	$1.28

Size Nemotron 3.5 Lightning 30B-A3B 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 3.5 Lightning 30B-A3B need?

The weights take 65.8 GB at the native BF16 precision (65.8 GB at BF16, 32.9 GB at FP8, 18.5 GB at INT4). Each concurrent stream adds 6.0 KB of KV cache per token: 0.0 GB at 4,096 tokens and 0.2 GB at 32,768 tokens.

What is the bare minimum to run Nemotron 3.5 Lightning 30B-A3B?

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 PRO 6000 Blackwell Max-Q Workstation Edition at $0.50 per hour.

How many H100s do you need to run Nemotron 3.5 Lightning 30B-A3B?

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 3.5 Lightning 30B-A3B cost per million tokens on an H100?

About $0.54 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 $0.29.

What precision does Nemotron 3.5 Lightning 30B-A3B 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. Nemotron 3 Nano geometry tuned for speed and released with speculative-decoding drafts. The checkpoint carries a multi-token-prediction layer, so it holds 32.9B tensors against the 31.6B the Hub reports. 6 attention layers keep a per-token cache; 23 Mamba-2 layers hold 49 MB of fixed FP32 state per sequence. NVIDIA also publishes official NVFP4 checkpoints.