Sarvam AI Sarvam

Sarvam 105B 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

106B

Active per token

10.3B

mixture of experts

KV cache per token

36 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Sarvam 105B

The strongest part each vendor offers for Sarvam 105B 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 2 x Ascend 950DT 4 x Gaudi 2 96GB 2 x TPU v7 192GB 8 x Inferentia2

Memory fills at: 2 x GB200 NVL72 GPU 186GB 120 users, 1 x Instinct MI355X OAM 53 users, 2 x Ascend 950DT 53 users, 4 x Gaudi 2 96GB 128 users, 2 x TPU v7 192GB 129 users, 8 x Inferentia2 22 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 285
Per hour: $5.99
Per million tokens: $5.83
Memory used: 77%

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: 229
Per hour: $2.00
Per million tokens: $2.42
Memory used: 58%

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: 468
Per hour: $2.00
Per million tokens: $1.19
Memory used: 68%

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	60 %	1.6k tok/s	77.7 tok/s	—	—	—	Specs →
2	Rubin SXM	1	90 %	1.4k tok/s	73.4 tok/s	—	—	—	Specs →
3	TPU 8t announced	2	60 %	~1.1k tok/s	41.4 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	90 %	~659 tok/s	28.7 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	90 %	625 tok/s	26.7 tok/s	$5.99	$2.66	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	90 %	624 tok/s	26.7 tok/s	$12.42	$5.53	—	Rent →
7	GB200 NVL72 GPU 186GB	2	70 %	1.2k tok/s	50.7 tok/s	$21.00	$4.92	—	Rent →
8	Instinct MI350X OAM	1	90 %	610 tok/s	26.7 tok/s	$6.16	$2.80	1.0 kW	Rent →
9	TPU v7 192GB	2	68 %	1.1k tok/s	46.8 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	60 %	~1.1k tok/s	77.7 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	99 %	605 tok/s	26.7 tok/s	$4.99	$2.29	1.4 kW	Rent →
12	B200 SXM 180GB	2	73 %	1.1k tok/s	48.8 tok/s	$8.18	$2.03	2.0 kW	Rent →
13	Ascend 970 announced	1	90 %	839 tok/s	48.0 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	2	68 %	1.1k tok/s	50.7 tok/s	—	—	1.4 kW	Specs →
15	Jalapeño announced	2	60 %	~1.5k tok/s	97.6 tok/s	—	—	1.4 kW	Specs →
16	Instinct MI325X OAM	2	51 %	795 tok/s	38.0 tok/s	$7.60	$2.65	2.0 kW	Rent →
17	Instinct MI300X 192GB	2	68 %	732 tok/s	33.6 tok/s	$4.78	$1.81	1.5 kW	Rent →
18	BR100	8	52 %	984 tok/s	36.3 tok/s	—	—	4.4 kW	Specs →
19	Ascend 960 announced	1	90 %	486 tok/s	32.0 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	2	93 %	624 tok/s	30.4 tok/s	$4.96	$2.21	1.4 kW	Rent →
21	H100 SXM5 80GB	4	82 %	926 tok/s	40.2 tok/s	$7.32	$2.20	2.8 kW	Rent →
22	TPU v6e 32GB 2 nodes	16	52 %	1.7k tok/s	65.6 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	4	51 %	868 tok/s	39.4 tok/s	—	—	2.4 kW	Specs →
24	H200 NVL 141GB	2	93 %	~585 tok/s	30.4 tok/s	$7.23	$3.43	1.2 kW	Rent →
25	H100 NVL 94GB	4	70 %	~974 tok/s	46.8 tok/s	$12.76	$3.64	1.6 kW	Rent →
26	Ascend 910C	4	51 %	841 tok/s	38.4 tok/s	—	—	2.4 kW	Specs →
27	H100 PCIe 80GB	4	82 %	589 tok/s	24.0 tok/s	$10.00	$4.72	1.4 kW	Rent →
28	H800 PCIe 80GB	4	82 %	589 tok/s	24.0 tok/s	—	—	1.4 kW	Specs →
29	Ascend 950DT	2	91 %	440 tok/s	25.4 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	4	68 %	488 tok/s	21.5 tok/s	$7.20	$4.10	2.4 kW	Rent →
31	Ascend 950PR	4	51 %	449 tok/s	19.2 tok/s	—	—	3.6 kW	Specs →
32	MTT S5000	4	82 %	~449 tok/s	19.2 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	4	68 %	448 tok/s	19.2 tok/s	$2.76	$1.71	2.4 kW	Rent →
34	Gaudi 3 128GB	4	51 %	1.1k tok/s	44.4 tok/s	—	—	3.6 kW	Specs →
35	TPU v5p 95GB	4	69 %	627 tok/s	33.2 tok/s	—	—	1.8 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	4	68 %	468 tok/s	21.5 tok/s	$2.00	$1.19	1.2 kW	Rent →
37	Cloud AI 100 Ultra	4	51 %	~184 tok/s	6.6 tok/s	—	—	0.6 kW	Specs →
38	Gaudi 2 96GB	4	68 %	569 tok/s	29.4 tok/s	$4.40	$2.15	2.4 kW	Rent →
39	Atlas 350	4	58 %	390 tok/s	16.8 tok/s	—	—	2.4 kW	Specs →
40	Data Center GPU Max 1100 48GB	8	69 %	667 tok/s	27.9 tok/s	—	—	2.4 kW	Specs →
41	GeForce RTX 5090 32GB 2 nodes	16	52 %	1.2k tok/s	71.7 tok/s	$19.04	$4.58	9.2 kW	Rent →
42	Ascend 910B	8	52 %	795 tok/s	36.3 tok/s	—	—	3.2 kW	Specs →
43	MI250X 128GB	4	51 %	634 tok/s	39.4 tok/s	—	—	2.2 kW	Specs →
44	RTX 6000 Ada 48GB	8	69 %	~534 tok/s	21.8 tok/s	$7.76	$4.03	2.4 kW	Rent →
45	Instinct MI250 128GB	4	51 %	616 tok/s	39.4 tok/s	—	—	2.0 kW	Specs →
46	L40S 48GB	8	69 %	492 tok/s	19.6 tok/s	$6.08	$3.43	2.8 kW	Rent →
47	A800 PCIe 80GB	4	82 %	433 tok/s	23.2 tok/s	—	—	1.2 kW	Specs →
48	A800 SXM4	4	82 %	447 tok/s	24.5 tok/s	—	—	1.6 kW	Specs →
49	A100 PCIe 80GB	4	82 %	433 tok/s	23.2 tok/s	$5.40	$3.46	1.2 kW	Rent →
50	A100 SXM4 80GB	4	82 %	447 tok/s	24.5 tok/s	$5.60	$3.48	1.6 kW	Rent →
51	A100 PCIe 40GB	8	83 %	716 tok/s	35.3 tok/s	—	—	2.0 kW	Specs →
52	A100 SXM4 40GB	8	83 %	716 tok/s	35.3 tok/s	$12.49	$4.84	3.2 kW	Rent →
53	GeForce RTX 5090 D V2 24GB 2 nodes	16	71 %	~1.1k tok/s	53.8 tok/s	—	—	9.2 kW	Specs →
54	RTX 5880 Ada 48GB	8	69 %	~498 tok/s	21.8 tok/s	—	—	2.3 kW	Specs →
55	TPU v4 32GB 2 nodes	16	52 %	1.0k tok/s	48.0 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB 2 nodes	16	52 %	~588 tok/s	23.1 tok/s	$13.28	$6.27	4.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	8	69 %	~609 tok/s	30.5 tok/s	$7.68	$3.50	2.4 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	4	91 %	~323 tok/s	16.1 tok/s	—	—	1.2 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition 2 nodes	16	52 %	704 tok/s	32.0 tok/s	—	—	2.6 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB 2 nodes	16	52 %	760 tok/s	35.9 tok/s	$11.52	$4.21	3.2 kW	Rent →
61	Radeon AI PRO R9700 32GB 2 nodes	16	52 %	591 tok/s	25.6 tok/s	—	—	4.8 kW	Specs →
62	Inferentia2 2 nodes	16	52 %	701 tok/s	32.8 tok/s	—	—	—	Specs →
63	Trainium 2 nodes	16	52 %	701 tok/s	32.8 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB 2 nodes	16	52 %	620 tok/s	49.2 tok/s	—	—	4.8 kW	Specs →
65	Arc Pro B70 32GB 2 nodes	16	52 %	~563 tok/s	24.3 tok/s	—	—	3.7 kW	Specs →
66	L40 48GB	8	69 %	404 tok/s	19.6 tok/s	$6.56	$4.51	2.4 kW	Rent →
67	Instinct MI210 PCIe	8	52 %	582 tok/s	37.2 tok/s	—	—	2.4 kW	Specs →
68	GeForce RTX 4090 24GB 2 nodes	16	71 %	758 tok/s	40.4 tok/s	$14.24	$5.22	7.2 kW	Rent →
69	A30 24GB 2 nodes	16	71 %	723 tok/s	37.3 tok/s	—	—	2.6 kW	Specs →
70	RTX 4500 Ada 24GB 2 nodes	16	71 %	~421 tok/s	17.3 tok/s	—	—	3.4 kW	Specs →
71	RTX A6000 48GB	8	69 %	354 tok/s	17.4 tok/s	$4.00	$3.13	2.4 kW	Rent →
72	A40 48GB	8	69 %	329 tok/s	15.8 tok/s	$4.72	$3.99	2.4 kW	Rent →
73	H20 141GB HBM3e	2	93 %	197 tok/s	25.4 tok/s	—	—	0.8 kW	Specs →
74	H20 96GB	4	68 %	373 tok/s	48.0 tok/s	—	—	1.6 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB 2 nodes	16	71 %	~564 tok/s	26.9 tok/s	$9.12	$4.49	2.3 kW	Rent →
76	BI-V100 Tiangai 100 2 nodes	16	52 %	791 tok/s	48.0 tok/s	—	—	4.0 kW	Specs →
77	Atlas 300I Duo 48GB	8	69 %	222 tok/s	9.3 tok/s	—	—	1.2 kW	Specs →
78	Atlas 300I Duo 96GB	4	68 %	117 tok/s	4.9 tok/s	—	—	0.6 kW	Specs →
79	RTX A5500 24GB 2 nodes	16	71 %	~596 tok/s	30.7 tok/s	—	—	3.7 kW	Specs →
80	Wormhole n300s 2 nodes	16	71 %	489 tok/s	23.1 tok/s	—	—	4.8 kW	Specs →
81	V100S PCIe 32GB 2 nodes	16	52 %	724 tok/s	45.4 tok/s	$14.08	$5.40	4.0 kW	Rent →
82	ATOM-Max	8	52 %	386 tok/s	23.2 tok/s	—	—	2.8 kW	Specs →
83	A10 24GB 2 nodes	16	71 %	495 tok/s	24.0 tok/s	$17.63	$9.90	2.4 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	8	69 %	346 tok/s	19.6 tok/s	—	—	2.4 kW	Specs →
85	Radeon PRO W7900 48GB	8	69 %	346 tok/s	19.6 tok/s	—	—	2.4 kW	Specs →
86	Radeon RX 7900 XTX 24GB 2 nodes	16	71 %	646 tok/s	38.4 tok/s	—	—	5.7 kW	Specs →
87	L4 24GB 2 nodes	16	71 %	299 tok/s	12.0 tok/s	$9.44	$8.78	1.2 kW	Rent →
88	L20 48GB	8	69 %	342 tok/s	19.6 tok/s	—	—	2.2 kW	Specs →
89	Radeon AI PRO R9600D 32GB 2 nodes	16	52 %	469 tok/s	25.6 tok/s	—	—	2.4 kW	Specs →
90	Arc Pro B65 32GB 2 nodes	16	52 %	~455 tok/s	24.3 tok/s	—	—	3.2 kW	Specs →
91	Arc Pro B60 24GB 2 nodes	16	71 %	~381 tok/s	18.3 tok/s	—	—	3.2 kW	Specs →
92	L2 24GB 2 nodes	16	71 %	283 tok/s	12.0 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB 2 nodes	16	71 %	~365 tok/s	17.3 tok/s	—	—	1.1 kW	Specs →
94	MLU370-X8	8	69 %	256 tok/s	13.9 tok/s	—	—	2.0 kW	Specs →
95	Radeon PRO W7800 48GB	8	69 %	298 tok/s	19.6 tok/s	—	—	2.1 kW	Specs →
96	Radeon PRO W7800 32GB 2 nodes	16	52 %	425 tok/s	23.1 tok/s	—	—	4.2 kW	Specs →
97	GeForce RTX 3090 24GB 2 nodes	16	71 %	475 tok/s	37.5 tok/s	$8.00	$4.67	5.6 kW	Rent →
98	A10G 24GB 2 nodes	16	71 %	386 tok/s	24.0 tok/s	$16.10	$11.57	4.8 kW	Rent →
99	Atlas 300V Pro	8	69 %	111 tok/s	4.6 tok/s	—	—	—	Specs →
100	Atlas 300I Pro 2 nodes	16	71 %	196 tok/s	8.2 tok/s	—	—	—	Specs →
101	Atlas 300V 2 nodes	16	71 %	178 tok/s	8.2 tok/s	—	—	1.2 kW	Specs →
102	GeForce RTX 3090 Ti 24GB 2 nodes	16	71 %	330 tok/s	40.4 tok/s	$7.36	$6.19	7.2 kW	Rent →
103	Radeon PRO W6800 32GB 2 nodes	16	52 %	247 tok/s	20.5 tok/s	—	—	4.0 kW	Specs →
104	RTX A5000 24GB 2 nodes	16	71 %	235 tok/s	30.7 tok/s	$4.32	$5.11	3.7 kW	Rent →
105	Radeon PRO V710 28GB 2 nodes	16	60 %	200 tok/s	17.9 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 468 tokens/s for $2.00/hr; the AMD pick is 2 x Instinct MI300X 192GB at about 732 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, 468 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, 732 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, 440 tok/s

Intel 7 parts fit

Best value: 4 x Gaudi 2 96GB, 569 tok/s for $4.40/hr

Fewest GPUs: 4 x Gaudi 3 128GB, 1.1k 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, 701 tok/s

Other 7 parts fit

Fewest GPUs: 4 x MTT S5000, 449 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	59k tok/s	$1,050	—	$4.92

Size Sarvam 105B 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 105B need?

The weights take 212 GB at the native BF16 precision (212 GB at BF16, 106 GB at FP8, 59.6 GB at INT4). Each concurrent stream adds 36 KB of KV cache per token: 0.2 GB at 4,096 tokens and 1.2 GB at 32,768 tokens.

What is the bare minimum to run Sarvam 105B?

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 Sarvam 105B?

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 Sarvam 105B cost per million tokens on an H100?

About $2.20 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 $1.19.

What precision does Sarvam 105B 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. MLA with 128 experts. The checkpoint is stored in FP32; weights here are the BF16 a server loads. Active parameters per the Sarvam model card.