NVIDIA Llama Nemotron

Llama 3.3 Nemotron Super 49B v1.5 hardware requirements

Fits on one Instinct MI355X OAM at BF16. For 1,000 tokens/s at 32k context you need 16 x H100 SXM5 80GB at $29.28/hr or 24 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $12.00/hr.

Parameters

49.9B

Active per token

All

dense model

KV cache per token

196 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Llama 3.3 Nemotron Super 49B v1.5

The strongest part each vendor offers for Llama 3.3 Nemotron Super 49B v1.5 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 950DT 2 x Gaudi 2 96GB 1 x TPU v7 192GB 4 x Inferentia2

Memory fills at: 1 x GB300 NVL72 GPU 288GB 26 users, 1 x Instinct MI355X OAM 26 users, 1 x Ascend 950DT 5 users, 2 x Gaudi 2 96GB 12 users, 1 x TPU v7 192GB 12 users, 4 x Inferentia2 2 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 58.9
Per hour: $5.99
Per million tokens: $28.27
Memory used: 36%

Rent on Daytona

Cheapest per hour

2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

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

Tokens/s: 25.0
Per hour: $1.00
Per million tokens: $11.13
Memory used: 55%

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: 188
Per hour: $2.00
Per million tokens: $2.95
Memory used: 85%

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	74 %	627 tok/s	56.3 tok/s	—	—	—	Specs →
2	Rubin SXM	2	56 %	1.0k tok/s	101 tok/s	—	—	—	Specs →
3	TPU 8t announced	2	75 %	~522 tok/s	30.0 tok/s	—	—	—	Specs →
4	TPU 8i announced	2	56 %	~531 tok/s	39.5 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	2	56 %	514 tok/s	36.8 tok/s	$11.98	$6.47	2.8 kW	Rent →
6	GB300 NVL72 GPU 288GB	2	56 %	512 tok/s	36.8 tok/s	$24.84	$13.47	—	Rent →
7	GB200 NVL72 GPU 186GB	2	87 %	512 tok/s	36.8 tok/s	$21.00	$11.39	—	Rent →
8	Instinct MI350X OAM	2	56 %	489 tok/s	36.8 tok/s	$12.32	$6.99	2.0 kW	Rent →
9	TPU v7 192GB	2	84 %	473 tok/s	33.9 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	74 %	~353 tok/s	56.3 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	2	61 %	482 tok/s	36.8 tok/s	$9.98	$5.75	2.8 kW	Rent →
12	B200 SXM 180GB	2	90 %	474 tok/s	35.4 tok/s	$8.18	$4.79	2.0 kW	Rent →
13	Ascend 970 announced	2	56 %	540 tok/s	66.2 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	2	84 %	412 tok/s	36.8 tok/s	—	—	1.4 kW	Specs →
15	Jalapeño announced	2	75 %	~485 tok/s	70.7 tok/s	—	—	1.4 kW	Specs →
16	Instinct MI325X OAM	2	63 %	308 tok/s	27.6 tok/s	$7.60	$6.84	2.0 kW	Rent →
17	Instinct MI300X 192GB	2	84 %	295 tok/s	24.3 tok/s	$4.78	$4.50	1.5 kW	Rent →
18	BR100	8	64 %	515 tok/s	26.3 tok/s	—	—	4.4 kW	Specs →
19	Ascend 960 announced	2	56 %	289 tok/s	44.1 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	4	57 %	451 tok/s	41.8 tok/s	$9.92	$6.11	2.8 kW	Rent →
21	H100 SXM5 80GB	8	51 %	743 tok/s	55.1 tok/s	$14.64	$5.47	5.6 kW	Rent →
22	TPU v6e 32GB 2 nodes	16	65 %	883 tok/s	47.5 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	4	63 %	353 tok/s	28.5 tok/s	—	—	2.4 kW	Specs →
24	H200 NVL 141GB	4	57 %	~402 tok/s	41.8 tok/s	$14.46	$9.99	2.4 kW	Rent →
25	H100 NVL 94GB	4	86 %	~376 tok/s	33.9 tok/s	$12.76	$9.43	1.6 kW	Rent →
26	Ascend 910C	4	63 %	340 tok/s	27.9 tok/s	—	—	2.4 kW	Specs →
27	H100 PCIe 80GB	8	51 %	508 tok/s	32.9 tok/s	$20.00	$10.93	2.8 kW	Rent →
28	H800 PCIe 80GB	8	51 %	508 tok/s	32.9 tok/s	—	—	2.8 kW	Specs →
29	Ascend 950DT	4	56 %	281 tok/s	34.8 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	4	85 %	204 tok/s	15.6 tok/s	$7.20	$9.78	2.4 kW	Rent →
31	Ascend 950PR	4	63 %	194 tok/s	13.9 tok/s	—	—	3.6 kW	Specs →
32	MTT S5000	8	51 %	~367 tok/s	26.3 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	4	85 %	194 tok/s	13.9 tok/s	$2.76	$3.95	2.4 kW	Rent →
34	Gaudi 3 128GB	4	63 %	544 tok/s	32.2 tok/s	—	—	3.6 kW	Specs →
35	TPU v5p 95GB	4	85 %	224 tok/s	24.1 tok/s	—	—	1.8 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	4	85 %	188 tok/s	15.6 tok/s	$2.00	$2.95	1.2 kW	Rent →
37	Cloud AI 100 Ultra	4	63 %	~101 tok/s	4.8 tok/s	—	—	0.6 kW	Specs →
38	Gaudi 2 96GB	4	85 %	207 tok/s	21.3 tok/s	$4.40	$5.91	2.4 kW	Rent →
39	Atlas 350	4	72 %	167 tok/s	12.2 tok/s	—	—	2.4 kW	Specs →
40	Data Center GPU Max 1100 48GB	8	86 %	296 tok/s	20.2 tok/s	—	—	2.4 kW	Specs →
41	GeForce RTX 5090 32GB 2 nodes	16	65 %	373 tok/s	52.0 tok/s	$19.04	$14.19	9.2 kW	Rent →
42	Ascend 910B	8	64 %	321 tok/s	26.3 tok/s	—	—	3.2 kW	Specs →
43	MI250X 128GB	4	63 %	204 tok/s	28.5 tok/s	—	—	2.2 kW	Specs →
44	RTX 6000 Ada 48GB	8	86 %	~244 tok/s	15.8 tok/s	$7.76	$8.82	2.4 kW	Rent →
45	Instinct MI250 128GB	4	63 %	196 tok/s	28.5 tok/s	—	—	2.0 kW	Specs →
46	L40S 48GB	8	86 %	231 tok/s	14.2 tok/s	$6.08	$7.31	2.8 kW	Rent →
47	A800 PCIe 80GB	8	51 %	290 tok/s	31.8 tok/s	—	—	2.4 kW	Specs →
48	A800 SXM4	8	51 %	294 tok/s	33.5 tok/s	—	—	3.2 kW	Specs →
49	A100 PCIe 80GB	8	51 %	290 tok/s	31.8 tok/s	$10.80	$10.35	2.4 kW	Rent →
50	A100 SXM4 80GB	8	51 %	294 tok/s	33.5 tok/s	$11.20	$10.58	3.2 kW	Rent →
51	A100 PCIe 40GB 2 nodes	16	52 %	478 tok/s	45.1 tok/s	—	—	4.0 kW	Specs →
52	A100 SXM4 40GB 2 nodes	16	52 %	478 tok/s	45.1 tok/s	$24.98	$14.51	6.4 kW	Rent →
53	GeForce RTX 5090 D V2 24GB 2 nodes	16	88 %	~440 tok/s	39.0 tok/s	—	—	9.2 kW	Specs →
54	RTX 5880 Ada 48GB	8	86 %	~210 tok/s	15.8 tok/s	—	—	2.3 kW	Specs →
55	TPU v4 32GB 2 nodes	16	65 %	403 tok/s	34.8 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB 2 nodes	16	65 %	~282 tok/s	16.7 tok/s	$13.28	$13.06	4.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	8	86 %	~227 tok/s	22.1 tok/s	$7.68	$9.38	2.4 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	8	57 %	~227 tok/s	22.1 tok/s	—	—	2.4 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition 2 nodes	16	65 %	286 tok/s	23.2 tok/s	—	—	2.6 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB 2 nodes	16	65 %	298 tok/s	26.0 tok/s	$11.52	$10.72	3.2 kW	Rent →
61	Radeon AI PRO R9700 32GB 2 nodes	16	65 %	252 tok/s	18.6 tok/s	—	—	4.8 kW	Specs →
62	Inferentia2 2 nodes	16	65 %	277 tok/s	23.8 tok/s	—	—	—	Specs →
63	Trainium 2 nodes	16	65 %	277 tok/s	23.8 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB 2 nodes	16	65 %	178 tok/s	35.7 tok/s	—	—	4.8 kW	Specs →
65	Arc Pro B70 32GB 2 nodes	16	65 %	~241 tok/s	17.6 tok/s	—	—	3.7 kW	Specs →
66	L40 48GB	8	86 %	155 tok/s	14.2 tok/s	$6.56	$11.76	2.4 kW	Rent →
67	Instinct MI210 PCIe	8	64 %	185 tok/s	26.9 tok/s	—	—	2.4 kW	Specs →
68	GeForce RTX 4090 24GB 2 nodes	16	88 %	270 tok/s	29.2 tok/s	$14.24	$14.68	7.2 kW	Rent →
69	A30 24GB 2 nodes	16	88 %	263 tok/s	27.1 tok/s	—	—	2.6 kW	Specs →
70	RTX 4500 Ada 24GB 2 nodes	16	88 %	~191 tok/s	12.5 tok/s	—	—	3.4 kW	Specs →
71	RTX A6000 48GB	8	86 %	134 tok/s	12.6 tok/s	$4.00	$8.28	2.4 kW	Rent →
72	A40 48GB	8	86 %	127 tok/s	11.4 tok/s	$4.72	$10.32	2.4 kW	Rent →
73	H20 141GB HBM3e	4	57 %	93.2 tok/s	34.8 tok/s	—	—	1.6 kW	Specs →
74	H20 96GB	4	85 %	93.2 tok/s	34.8 tok/s	—	—	1.6 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB 2 nodes	16	88 %	~219 tok/s	19.5 tok/s	$9.12	$11.57	2.3 kW	Rent →
76	BI-V100 Tiangai 100 2 nodes	16	65 %	259 tok/s	34.8 tok/s	—	—	4.0 kW	Specs →
77	Atlas 300I Duo 48GB	8	86 %	98.4 tok/s	6.7 tok/s	—	—	1.2 kW	Specs →
78	Atlas 300I Duo 96GB	4	85 %	52.1 tok/s	3.6 tok/s	—	—	0.6 kW	Specs →
79	RTX A5500 24GB 2 nodes	16	88 %	~217 tok/s	22.3 tok/s	—	—	3.7 kW	Specs →
80	Wormhole n300s 2 nodes	16	88 %	192 tok/s	16.7 tok/s	—	—	4.8 kW	Specs →
81	V100S PCIe 32GB 2 nodes	16	65 %	232 tok/s	32.9 tok/s	$14.08	$16.84	4.0 kW	Rent →
82	ATOM-Max	8	64 %	127 tok/s	16.8 tok/s	—	—	2.8 kW	Specs →
83	A10 24GB 2 nodes	16	88 %	189 tok/s	17.4 tok/s	$17.63	$25.88	2.4 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	8	86 %	118 tok/s	14.2 tok/s	—	—	2.4 kW	Specs →
85	Radeon PRO W7900 48GB	8	86 %	118 tok/s	14.2 tok/s	—	—	2.4 kW	Specs →
86	Radeon RX 7900 XTX 24GB 2 nodes	16	88 %	214 tok/s	27.9 tok/s	—	—	5.7 kW	Specs →
87	L4 24GB 2 nodes	16	88 %	139 tok/s	8.7 tok/s	$9.44	$18.87	1.2 kW	Rent →
88	L20 48GB	8	86 %	116 tok/s	14.2 tok/s	—	—	2.2 kW	Specs →
89	Radeon AI PRO R9600D 32GB 2 nodes	16	65 %	164 tok/s	18.6 tok/s	—	—	2.4 kW	Specs →
90	Arc Pro B65 32GB 2 nodes	16	65 %	~161 tok/s	17.6 tok/s	—	—	3.2 kW	Specs →
91	Arc Pro B60 24GB 2 nodes	16	88 %	~147 tok/s	13.2 tok/s	—	—	3.2 kW	Specs →
92	L2 24GB 2 nodes	16	88 %	123 tok/s	8.7 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB 2 nodes	16	88 %	~142 tok/s	12.5 tok/s	—	—	1.1 kW	Specs →
94	MLU370-X8	8	86 %	90.0 tok/s	10.1 tok/s	—	—	2.0 kW	Specs →
95	Radeon PRO W7800 48GB	8	86 %	93.3 tok/s	14.2 tok/s	—	—	2.1 kW	Specs →
96	Radeon PRO W7800 32GB 2 nodes	16	65 %	149 tok/s	16.7 tok/s	—	—	4.2 kW	Specs →
97	GeForce RTX 3090 24GB 2 nodes	16	88 %	137 tok/s	27.2 tok/s	$8.00	$16.21	5.6 kW	Rent →
98	A10G 24GB 2 nodes	16	88 %	125 tok/s	17.4 tok/s	$16.10	$35.90	4.8 kW	Rent →
99	Atlas 300V Pro	8	86 %	49.3 tok/s	3.4 tok/s	—	—	—	Specs →
100	Atlas 300I Pro 2 nodes	16	88 %	87.0 tok/s	5.9 tok/s	—	—	—	Specs →
101	Atlas 300V 2 nodes	16	88 %	71.5 tok/s	5.9 tok/s	—	—	1.2 kW	Specs →
102	GeForce RTX 3090 Ti 24GB 2 nodes	16	88 %	83.5 tok/s	29.2 tok/s	$7.36	$24.49	7.2 kW	Rent →
103	Radeon PRO W6800 32GB 2 nodes	16	65 %	69.8 tok/s	14.9 tok/s	—	—	4.0 kW	Specs →
104	RTX A5000 24GB 2 nodes	16	88 %	58.5 tok/s	22.3 tok/s	$4.32	$20.52	3.7 kW	Rent →
105	Radeon PRO V710 28GB 2 nodes	16	75 %	55.0 tok/s	13.0 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 188 tokens/s for $2.00/hr; the AMD pick is 2 x Instinct MI300X 192GB at about 295 tokens/s for $4.78/hr. Per rental dollar NVIDIA delivers 1.52x the tokens of AMD here.

NVIDIA 51 parts fit

Best value: 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 188 tok/s for $2.00/hr

Fewest GPUs: 2 x Rubin SXM, 1.0k tok/s

AMD 18 parts fit

Best value: 2 x Instinct MI300X 192GB, 295 tok/s for $4.78/hr

Fewest GPUs: 1 x Instinct MI455X, 627 tok/s

Huawei 10 parts fit

Fewest GPUs: 4 x Ascend 910C, 340 tok/s

Intel 7 parts fit

Best value: 4 x Gaudi 2 96GB, 207 tok/s for $4.40/hr

Fewest GPUs: 4 x Gaudi 3 128GB, 544 tok/s

Google 4 parts fit

Fewest GPUs: 2 x TPU v7 192GB, 473 tok/s

AWS 2 parts fit

Fewest GPUs: 16 x Inferentia2, 277 tok/s

Other 7 parts fit

Fewest GPUs: 4 x Cloud AI 100 Ultra, 101 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	250 x TP4	0	113k tok/s	$2,480	700 kW	$6.11
B	100 x GB200 NVL72 GPU 186GB	50 x TP2	0	26k tok/s	$1,050	—	$11.39

Size Llama 3.3 Nemotron Super 49B v1.5 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 3.3 Nemotron Super 49B v1.5 need?

The weights take 99.7 GB at the native BF16 precision (99.7 GB at BF16, 49.9 GB at FP8, 28.1 GB at INT4). Each concurrent stream adds 196 KB of KV cache per token: 0.8 GB at 4,096 tokens and 6.6 GB at 32,768 tokens.

What is the bare minimum to run Llama 3.3 Nemotron Super 49B v1.5?

1 x Instinct MI355X OAM at BF16 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.00 per hour.

How many H100s do you need to run Llama 3.3 Nemotron Super 49B v1.5?

2 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 8 H100s, and 1,000 tokens per second takes 16 H100s across 2 replicas.

How much does Llama 3.3 Nemotron Super 49B v1.5 cost per million tokens on an H100?

About $5.47 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 $2.95.

What precision does Llama 3.3 Nemotron Super 49B v1.5 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 3.3 70B pruned by neural architecture search: 49 of 80 layers keep attention, so the cache is 196 KB per token rather than 320 KB. NVIDIA tested it on 2 A100 80GB.