Sarvam AI Sarvam

Sarvam 30B 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

32.2B

Active per token

2.4B

mixture of experts

KV cache per token

19 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Sarvam 30B

The strongest part each vendor offers for Sarvam 30B 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 1 x Gaudi 2 96GB 1 x TPU v7 192GB 4 x Inferentia2

Memory fills at: 1 x GB300 NVL72 GPU 288GB 333 users, 1 x Instinct MI355X OAM 333 users, 1 x Ascend 950DT 115 users, 1 x Gaudi 2 96GB 42 users, 1 x TPU v7 192GB 188 users, 4 x Inferentia2 86 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 1.2k
Per hour: $5.99
Per million tokens: $1.37
Memory used: 23%

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: 271
Per hour: $0.50
Per million tokens: $0.51
Memory used: 70%

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: 448
Per hour: $0.50
Per million tokens: $0.31
Memory used: 92%

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	20 %	5.6k tok/s	245 tok/s	—	—	—	Specs →
2	Rubin SXM	1	31 %	5.1k tok/s	232 tok/s	—	—	—	Specs →
3	TPU 8t announced	1	41 %	~1.9k tok/s	68.7 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	31 %	~2.2k tok/s	90.5 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	31 %	2.1k tok/s	84.2 tok/s	$5.99	$0.78	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	31 %	2.1k tok/s	84.2 tok/s	$12.42	$1.63	—	Rent →
7	GB200 NVL72 GPU 186GB	1	47 %	2.1k tok/s	84.2 tok/s	$10.50	$1.38	—	Rent →
8	Instinct MI350X OAM	1	31 %	2.1k tok/s	84.2 tok/s	$6.16	$0.82	1.0 kW	Rent →
9	TPU v7 192GB	1	46 %	2.0k tok/s	77.7 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	20 %	~4.2k tok/s	245 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	34 %	2.1k tok/s	84.2 tok/s	$4.99	$0.67	1.4 kW	Rent →
12	B200 SXM 180GB	1	49 %	2.0k tok/s	81.0 tok/s	$4.09	$0.57	1.0 kW	Rent →
13	Ascend 970 announced	1	31 %	3.0k tok/s	152 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	1	46 %	1.9k tok/s	84.2 tok/s	—	—	0.7 kW	Specs →
15	Jalapeño announced	1	41 %	~2.9k tok/s	162 tok/s	—	—	0.7 kW	Specs →
16	Instinct MI325X OAM	1	34 %	1.5k tok/s	63.1 tok/s	$3.80	$0.73	1.0 kW	Rent →
17	Instinct MI300X 192GB	1	46 %	1.3k tok/s	55.8 tok/s	$2.39	$0.50	0.8 kW	Rent →
18	BR100	2	70 %	904 tok/s	32.0 tok/s	—	—	1.1 kW	Specs →
19	Ascend 960 announced	1	31 %	1.8k tok/s	101 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	1	63 %	1.1k tok/s	50.5 tok/s	$2.48	$0.60	0.7 kW	Rent →
21	H100 SXM5 80GB	2	56 %	1.7k tok/s	67.0 tok/s	$3.66	$0.61	1.4 kW	Rent →
22	TPU v6e 32GB	4	71 %	1.7k tok/s	62.1 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	1	69 %	829 tok/s	34.5 tok/s	—	—	0.6 kW	Specs →
24	H200 NVL 141GB	1	63 %	~1.1k tok/s	50.5 tok/s	$3.61	$0.92	0.6 kW	Rent →
25	H100 NVL 94GB	1	94 %	~941 tok/s	41.0 tok/s	$3.19	$0.94	0.4 kW	Rent →
26	Ascend 910C	1	69 %	804 tok/s	33.7 tok/s	—	—	0.6 kW	Specs →
27	H100 PCIe 80GB	2	56 %	1.0k tok/s	40.0 tok/s	$5.00	$1.33	0.7 kW	Rent →
28	H800 PCIe 80GB	2	56 %	1.0k tok/s	40.0 tok/s	—	—	0.7 kW	Specs →
29	Ascend 950DT	1	61 %	831 tok/s	42.1 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	1	92 %	464 tok/s	18.9 tok/s	$1.80	$1.08	0.6 kW	Rent →
31	Ascend 950PR	1	69 %	424 tok/s	16.8 tok/s	—	—	0.9 kW	Specs →
32	MTT S5000	2	56 %	~805 tok/s	32.0 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	1	92 %	423 tok/s	16.8 tok/s	$0.69	$0.45	0.6 kW	Rent →
34	Gaudi 3 128GB	1	69 %	1.0k tok/s	38.9 tok/s	—	—	0.9 kW	Specs →
35	TPU v5p 95GB	1	93 %	616 tok/s	29.1 tok/s	—	—	0.5 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	1	92 %	448 tok/s	18.9 tok/s	$0.50	$0.31	0.3 kW	Rent →
37	Cloud AI 100 Ultra	1	69 %	~167 tok/s	5.8 tok/s	—	—	0.1 kW	Specs →
38	Gaudi 2 96GB	1	92 %	557 tok/s	25.8 tok/s	$1.10	$0.55	0.6 kW	Rent →
39	Atlas 350	1	79 %	368 tok/s	14.7 tok/s	—	—	0.6 kW	Specs →
40	Data Center GPU Max 1100 48GB	2	93 %	630 tok/s	24.6 tok/s	—	—	0.6 kW	Specs →
41	GeForce RTX 5090 32GB	4	71 %	1.3k tok/s	67.9 tok/s	$4.76	$1.05	2.3 kW	Rent →
42	Ascend 910B	2	70 %	764 tok/s	32.0 tok/s	—	—	0.8 kW	Specs →
43	MI250X 128GB	1	69 %	639 tok/s	34.5 tok/s	—	—	0.6 kW	Specs →
44	RTX 6000 Ada 48GB	2	93 %	~502 tok/s	19.2 tok/s	$1.94	$1.07	0.6 kW	Rent →
45	Instinct MI250 128GB	1	69 %	624 tok/s	34.5 tok/s	—	—	0.5 kW	Specs →
46	L40S 48GB	2	93 %	460 tok/s	17.3 tok/s	$1.52	$0.92	0.7 kW	Rent →
47	A800 PCIe 80GB	2	56 %	811 tok/s	38.7 tok/s	—	—	0.6 kW	Specs →
48	A800 SXM4	2	56 %	839 tok/s	40.8 tok/s	—	—	0.8 kW	Specs →
49	A100 PCIe 80GB	2	56 %	811 tok/s	38.7 tok/s	$2.70	$0.92	0.6 kW	Rent →
50	A100 SXM4 80GB	2	56 %	839 tok/s	40.8 tok/s	$2.80	$0.93	0.8 kW	Rent →
51	A100 PCIe 40GB	4	56 %	1.3k tok/s	58.9 tok/s	—	—	1.0 kW	Specs →
52	A100 SXM4 40GB	4	56 %	1.3k tok/s	58.9 tok/s	$6.24	$1.31	1.6 kW	Rent →
53	GeForce RTX 5090 D V2 24GB	4	96 %	~1.2k tok/s	50.9 tok/s	—	—	2.3 kW	Specs →
54	RTX 5880 Ada 48GB	2	93 %	~475 tok/s	19.2 tok/s	—	—	0.6 kW	Specs →
55	TPU v4 32GB	4	71 %	1.1k tok/s	45.5 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB	4	71 %	~588 tok/s	21.8 tok/s	$3.32	$1.57	1.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	2	93 %	~596 tok/s	26.9 tok/s	$1.92	$0.89	0.6 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	2	62 %	~596 tok/s	26.9 tok/s	—	—	0.6 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition	4	71 %	727 tok/s	30.3 tok/s	—	—	0.7 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB	4	71 %	790 tok/s	33.9 tok/s	$2.88	$1.01	0.8 kW	Rent →
61	Radeon AI PRO R9700 32GB	4	71 %	604 tok/s	24.2 tok/s	—	—	1.2 kW	Specs →
62	Inferentia2	4	71 %	728 tok/s	31.1 tok/s	—	—	—	Specs →
63	Trainium	4	71 %	728 tok/s	31.1 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB	4	71 %	699 tok/s	46.6 tok/s	—	—	1.2 kW	Specs →
65	Arc Pro B70 32GB	4	71 %	~575 tok/s	23.0 tok/s	—	—	0.9 kW	Specs →
66	L40 48GB	2	93 %	393 tok/s	17.3 tok/s	$1.64	$1.16	0.6 kW	Rent →
67	Instinct MI210 PCIe	2	70 %	593 tok/s	32.8 tok/s	—	—	0.6 kW	Specs →
68	GeForce RTX 4090 24GB	4	96 %	805 tok/s	38.2 tok/s	$3.56	$1.23	1.8 kW	Rent →
69	A30 24GB	4	96 %	764 tok/s	35.3 tok/s	—	—	0.7 kW	Specs →
70	RTX 4500 Ada 24GB	4	96 %	~425 tok/s	16.4 tok/s	—	—	0.8 kW	Specs →
71	RTX A6000 48GB	2	93 %	346 tok/s	15.4 tok/s	$1.00	$0.80	0.6 kW	Rent →
72	A40 48GB	2	93 %	319 tok/s	13.9 tok/s	$1.18	$1.03	0.6 kW	Rent →
73	H20 141GB HBM3e	1	63 %	409 tok/s	42.1 tok/s	—	—	0.4 kW	Specs →
74	H20 96GB	1	92 %	409 tok/s	42.1 tok/s	—	—	0.4 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB	4	96 %	~587 tok/s	25.5 tok/s	$2.28	$1.08	0.6 kW	Rent →
76	BI-V100 Tiangai 100	4	71 %	859 tok/s	45.5 tok/s	—	—	1.0 kW	Specs →
77	Atlas 300I Duo 48GB	2	93 %	209 tok/s	8.2 tok/s	—	—	0.3 kW	Specs →
78	Atlas 300I Duo 96GB	1	92 %	110 tok/s	4.3 tok/s	—	—	0.1 kW	Specs →
79	RTX A5500 24GB	4	96 %	~630 tok/s	29.1 tok/s	—	—	0.9 kW	Specs →
80	Wormhole n300s	4	96 %	509 tok/s	21.8 tok/s	—	—	1.2 kW	Specs →
81	V100S PCIe 32GB	4	71 %	790 tok/s	43.0 tok/s	$3.52	$1.24	1.0 kW	Rent →
82	ATOM-Max	2	70 %	390 tok/s	20.5 tok/s	—	—	0.7 kW	Specs →
83	A10 24GB	4	96 %	517 tok/s	22.7 tok/s	$4.41	$2.37	0.6 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	2	93 %	346 tok/s	17.3 tok/s	—	—	0.6 kW	Specs →
85	Radeon PRO W7900 48GB	2	93 %	346 tok/s	17.3 tok/s	—	—	0.6 kW	Specs →
86	Radeon RX 7900 XTX 24GB	4	96 %	700 tok/s	36.4 tok/s	—	—	1.4 kW	Specs →
87	L4 24GB	4	96 %	300 tok/s	11.4 tok/s	$2.36	$2.18	0.3 kW	Rent →
88	L20 48GB	2	93 %	343 tok/s	17.3 tok/s	—	—	0.6 kW	Specs →
89	Radeon AI PRO R9600D 32GB	4	71 %	501 tok/s	24.2 tok/s	—	—	0.6 kW	Specs →
90	Arc Pro B65 32GB	4	71 %	~484 tok/s	23.0 tok/s	—	—	0.8 kW	Specs →
91	Arc Pro B60 24GB	4	96 %	~397 tok/s	17.3 tok/s	—	—	0.8 kW	Specs →
92	L2 24GB	4	96 %	288 tok/s	11.4 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB	4	96 %	~379 tok/s	16.4 tok/s	—	—	0.3 kW	Specs →
94	MLU370-X8	2	93 %	255 tok/s	12.3 tok/s	—	—	0.5 kW	Specs →
95	Radeon PRO W7800 48GB	2	93 %	305 tok/s	17.3 tok/s	—	—	0.5 kW	Specs →
96	Radeon PRO W7800 32GB	4	71 %	453 tok/s	21.8 tok/s	—	—	1.0 kW	Specs →
97	GeForce RTX 3090 24GB	4	96 %	536 tok/s	35.5 tok/s	$2.00	$1.04	1.4 kW	Rent →
98	A10G 24GB	4	96 %	421 tok/s	22.7 tok/s	$4.02	$2.65	1.2 kW	Rent →
99	Atlas 300V Pro	2	93 %	105 tok/s	4.1 tok/s	—	—	—	Specs →
100	Atlas 300I Pro	4	96 %	199 tok/s	7.8 tok/s	—	—	—	Specs →
101	Atlas 300V	4	96 %	184 tok/s	7.8 tok/s	—	—	0.3 kW	Specs →
102	GeForce RTX 3090 Ti 24GB	4	96 %	389 tok/s	38.2 tok/s	$1.84	$1.31	1.8 kW	Rent →
103	Radeon PRO W6800 32GB	4	71 %	280 tok/s	19.4 tok/s	—	—	1.0 kW	Specs →
104	RTX A5000 24GB	4	96 %	278 tok/s	29.1 tok/s	$1.08	$1.08	0.9 kW	Rent →
105	Radeon PRO V710 28GB	4	81 %	229 tok/s	17.0 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 448 tokens/s for $0.50/hr; the AMD pick is 1 x Instinct MI300X 192GB at about 1.3k 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, 448 tok/s for $0.50/hr

Fewest GPUs: 1 x Rubin SXM, 5.1k tok/s

AMD 18 parts fit

Best value: 1 x Instinct MI300X 192GB, 1.3k tok/s for $2.39/hr

Fewest GPUs: 1 x Instinct MI455X, 5.6k tok/s

Huawei 10 parts fit

Fewest GPUs: 1 x Ascend 950DT, 831 tok/s

Intel 7 parts fit

Best value: 1 x Gaudi 2 96GB, 557 tok/s for $1.10/hr

Fewest GPUs: 1 x Gaudi 3 128GB, 1.0k tok/s

Google 4 parts fit

Fewest GPUs: 1 x TPU v7 192GB, 2.0k tok/s

AWS 2 parts fit

Fewest GPUs: 4 x Inferentia2, 728 tok/s

Other 7 parts fit

Fewest GPUs: 1 x Cloud AI 100 Ultra, 167 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	1.15M tok/s	$2,480	700 kW	$0.60
B	100 x GB200 NVL72 GPU 186GB	100 x TP1	0	212k tok/s	$1,050	—	$1.38

Size Sarvam 30B 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 Sarvam 30B need?

The weights take 64.3 GB at the native BF16 precision (64.3 GB at BF16, 32.2 GB at FP8, 18.1 GB at INT4). Each concurrent stream adds 19 KB of KV cache per token: 0.1 GB at 4,096 tokens and 0.6 GB at 32,768 tokens.

What is the bare minimum to run Sarvam 30B?

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 Sarvam 30B?

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

How much does Sarvam 30B cost per million tokens on an H100?

About $0.61 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 $0.31.

What precision does Sarvam 30B 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. The checkpoint is stored in FP32; weights here are the BF16 a server loads. Active parameters per the Sarvam card, which counts them without embeddings.