OpenAI gpt-oss

gpt-oss-20b hardware requirements

Fits on one GB300 NVL72 GPU 288GB at INT4 / FP4. For 1,000 tokens/s at 32k context you need 1 x H100 SXM5 80GB at $1.83/hr or 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.00/hr.

Parameters

20.9B

Active per token

3.6B

mixture of experts

KV cache per token

48 KB

16-bit cache

Native precision

MXFP4

Rent Model card

Top options for gpt-oss-20b

The strongest part each vendor offers for gpt-oss-20b at INT4 / FP4 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 3 128GB 1 x TPU v7 192GB 1 x Inferentia2

Memory fills at: 1 x GB300 NVL72 GPU 288GB 328 users, 1 x Instinct MI355X OAM 328 users, 1 x Ascend 950DT 156 users, 1 x Gaudi 3 128GB 137 users, 1 x TPU v7 192GB 213 users, 1 x Inferentia2 22 users. Roofline estimates at INT4 / FP4.

Bare minimum

Fewest GPUs

1 x GB300 NVL72 GPU 288GB

INT4 / FP4, 4,096 tokens, one stream

Tokens/s: 2.8k
Per hour: $12.42
Per million tokens: $1.23
Memory used: 4%

Rent on DataCrunch

Cheapest per hour

1 x RTX A5000 24GB

INT4 / FP4, 4,096 tokens, one stream, live price

Tokens/s: 192
Per hour: $0.27
Per million tokens: $0.39
Memory used: 54%

Rent on RunPod

Cheapest per million tokens

1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

INT4 / FP4, 32,768 tokens, 32 streams

Tokens/s: 920
Per hour: $0.50
Per million tokens: $0.15
Memory used: 41%

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; 68 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	9 %	13k tok/s	466 tok/s	—	—	—	Specs →
2	Rubin SXM	1	14 %	12k tok/s	440 tok/s	—	—	—	Specs →
3	TPU 8t announced	1	18 %	~3.7k tok/s	131 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	14 %	~4.6k tok/s	172 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	14 %	4.3k tok/s	160 tok/s	$5.99	$0.39	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	14 %	4.5k tok/s	160 tok/s	$12.42	$0.76	—	Rent →
7	GB200 NVL72 GPU 186GB	1	21 %	4.3k tok/s	160 tok/s	$10.50	$0.68	—	Rent →
8	Instinct MI350X OAM	1	14 %	4.2k tok/s	160 tok/s	$6.16	$0.40	1.0 kW	Rent →
9	TPU v7 192GB	1	20 %	~2.7k tok/s	148 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	9 %	~9.2k tok/s	466 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	15 %	4.5k tok/s	160 tok/s	$4.99	$0.31	1.4 kW	Rent →
12	B200 SXM 180GB	1	22 %	4.1k tok/s	154 tok/s	$4.09	$0.28	1.0 kW	Rent →
13	Ascend 970 announced	1	14 %	6.4k tok/s	288 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	1	20 %	4.0k tok/s	160 tok/s	—	—	0.7 kW	Specs →
15	Jalapeño announced	1	18 %	~7.7k tok/s	308 tok/s	—	—	0.7 kW	Specs →
16	Instinct MI325X OAM	1	15 %	~1.8k tok/s	120 tok/s	$3.80	$0.58	1.0 kW	Rent →
17	Instinct MI300X 192GB	1	20 %	~1.7k tok/s	106 tok/s	$2.39	$0.39	0.8 kW	Rent →
18	BR100	1	62 %	~743 tok/s	32.0 tok/s	—	—	0.6 kW	Specs →
19	Ascend 960 announced	1	14 %	3.9k tok/s	192 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	1	28 %	~1.4k tok/s	96.1 tok/s	$2.48	$0.49	0.7 kW	Rent →
21	H100 SXM5 80GB	1	49 %	~1.2k tok/s	67.0 tok/s	$1.83	$0.43	0.7 kW	Rent →
22	TPU v6e 32GB	2	63 %	~1.4k tok/s	62.3 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	1	31 %	~1.1k tok/s	65.6 tok/s	—	—	0.6 kW	Specs →
24	H200 NVL 141GB	1	28 %	~1.3k tok/s	96.1 tok/s	$3.61	$0.78	0.6 kW	Rent →
25	H100 NVL 94GB	1	42 %	~1.2k tok/s	78.0 tok/s	$3.19	$0.76	0.4 kW	Rent →
26	Ascend 910C	1	31 %	~1.0k tok/s	64.0 tok/s	—	—	0.6 kW	Specs →
27	H100 PCIe 80GB	1	49 %	~781 tok/s	40.0 tok/s	$2.50	$0.89	0.3 kW	Rent →
28	H800 PCIe 80GB	1	49 %	~781 tok/s	40.0 tok/s	—	—	0.3 kW	Specs →
29	Ascend 950DT	1	27 %	1.7k tok/s	80.0 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	1	41 %	944 tok/s	35.9 tok/s	$1.80	$0.53	0.6 kW	Rent →
31	Ascend 950PR	1	31 %	858 tok/s	32.0 tok/s	—	—	0.9 kW	Specs →
32	MTT S5000	1	49 %	~577 tok/s	32.0 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	1	41 %	857 tok/s	32.0 tok/s	$0.69	$0.22	0.6 kW	Rent →
34	Gaudi 3 128GB	1	31 %	~1.5k tok/s	74.0 tok/s	—	—	0.9 kW	Specs →
35	TPU v5p 95GB	1	41 %	~720 tok/s	55.3 tok/s	—	—	0.5 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	1	41 %	920 tok/s	35.9 tok/s	$0.50	$0.15	0.3 kW	Rent →
37	Cloud AI 100 Ultra	1	31 %	~269 tok/s	11.0 tok/s	—	—	0.1 kW	Specs →
38	Gaudi 2 96GB	1	41 %	~661 tok/s	49.0 tok/s	—	—	0.6 kW	Specs →
39	Atlas 350	1	35 %	737 tok/s	28.0 tok/s	—	—	0.6 kW	Specs →
40	Data Center GPU Max 1100 48GB	1	83 %	~460 tok/s	24.6 tok/s	—	—	0.3 kW	Specs →
41	GeForce RTX 5090 32GB	2	63 %	1.7k tok/s	68.1 tok/s	$2.38	$0.38	1.1 kW	Rent →
42	Ascend 910B	1	62 %	~521 tok/s	32.0 tok/s	—	—	0.4 kW	Specs →
43	MI250X 128GB	1	31 %	~683 tok/s	65.6 tok/s	—	—	0.6 kW	Specs →
44	RTX 6000 Ada 48GB	1	83 %	~375 tok/s	19.2 tok/s	$0.99	$0.73	0.3 kW	Rent →
45	Instinct MI250 128GB	1	31 %	~658 tok/s	65.6 tok/s	—	—	0.5 kW	Specs →
46	L40S 48GB	1	83 %	~351 tok/s	17.3 tok/s	$0.76	$0.60	0.3 kW	Rent →
47	A800 PCIe 80GB	1	49 %	~495 tok/s	38.7 tok/s	—	—	0.3 kW	Specs →
48	A800 SXM4	1	49 %	~506 tok/s	40.8 tok/s	—	—	0.4 kW	Specs →
49	A100 PCIe 80GB	1	49 %	~495 tok/s	38.7 tok/s	$1.35	$0.76	0.3 kW	Rent →
50	A100 SXM4 80GB	1	49 %	~506 tok/s	40.8 tok/s	$1.35	$0.74	0.4 kW	Rent →
51	A100 PCIe 40GB	2	50 %	~857 tok/s	59.1 tok/s	—	—	0.5 kW	Specs →
52	A100 SXM4 40GB	2	50 %	~857 tok/s	59.1 tok/s	$3.12	$1.01	0.8 kW	Rent →
53	GeForce RTX 5090 D V2 24GB	2	85 %	~1.3k tok/s	51.1 tok/s	—	—	1.1 kW	Specs →
54	RTX 5880 Ada 48GB	1	83 %	~334 tok/s	19.2 tok/s	—	—	0.3 kW	Specs →
55	TPU v4 32GB	2	63 %	~710 tok/s	45.6 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB	2	63 %	~457 tok/s	21.9 tok/s	$1.66	$1.01	0.5 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	1	83 %	~654 tok/s	26.9 tok/s	$0.96	$0.41	0.3 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	1	55 %	~654 tok/s	26.9 tok/s	—	—	0.3 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition	2	63 %	784 tok/s	30.4 tok/s	—	—	0.3 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB	2	63 %	860 tok/s	34.1 tok/s	$1.44	$0.47	0.4 kW	Rent →
61	Radeon AI PRO R9700 32GB	2	63 %	~430 tok/s	24.3 tok/s	—	—	0.6 kW	Specs →
62	Inferentia2	2	63 %	~488 tok/s	31.2 tok/s	—	—	—	Specs →
63	Trainium	2	63 %	~488 tok/s	31.2 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB	2	63 %	~354 tok/s	46.7 tok/s	—	—	0.6 kW	Specs →
65	Arc Pro B70 32GB	2	63 %	~411 tok/s	23.1 tok/s	—	—	0.5 kW	Specs →
66	L40 48GB	1	83 %	~257 tok/s	17.3 tok/s	$0.82	$0.89	0.3 kW	Rent →
67	Instinct MI210 PCIe	1	62 %	~329 tok/s	32.8 tok/s	—	—	0.3 kW	Specs →
68	GeForce RTX 4090 24GB	2	85 %	~495 tok/s	38.3 tok/s	$0.80	$0.45	0.9 kW	Rent →
69	A30 24GB	2	85 %	~479 tok/s	35.5 tok/s	—	—	0.3 kW	Specs →
70	RTX 4500 Ada 24GB	2	85 %	~317 tok/s	16.4 tok/s	—	—	0.4 kW	Specs →
71	RTX A6000 48GB	1	83 %	~223 tok/s	15.4 tok/s	$0.50	$0.62	0.3 kW	Rent →
72	A40 48GB	1	83 %	~210 tok/s	13.9 tok/s	$0.59	$0.78	0.3 kW	Rent →
73	H20 141GB HBM3e	1	28 %	~340 tok/s	80.0 tok/s	—	—	0.4 kW	Specs →
74	H20 96GB	1	41 %	~340 tok/s	80.0 tok/s	—	—	0.4 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB	2	85 %	~641 tok/s	25.5 tok/s	$1.14	$0.49	0.3 kW	Rent →
76	BI-V100 Tiangai 100	2	63 %	~491 tok/s	45.6 tok/s	—	—	0.5 kW	Specs →
77	Atlas 300I Duo 48GB	1	83 %	~153 tok/s	8.2 tok/s	—	—	0.1 kW	Specs →
78	Atlas 300I Duo 96GB	1	41 %	~153 tok/s	8.2 tok/s	—	—	0.1 kW	Specs →
79	RTX A5500 24GB	2	85 %	~395 tok/s	29.2 tok/s	—	—	0.5 kW	Specs →
80	Wormhole n300s	2	85 %	~339 tok/s	21.9 tok/s	—	—	0.6 kW	Specs →
81	V100S PCIe 32GB	2	63 %	~443 tok/s	43.1 tok/s	$1.76	$1.10	0.5 kW	Rent →
82	ATOM-Max	1	62 %	~223 tok/s	20.5 tok/s	—	—	0.3 kW	Specs →
83	A10 24GB	2	85 %	~338 tok/s	22.8 tok/s	$2.20	$1.81	0.3 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	1	83 %	~205 tok/s	17.3 tok/s	—	—	0.3 kW	Specs →
85	Radeon PRO W7900 48GB	1	83 %	~205 tok/s	17.3 tok/s	—	—	0.3 kW	Specs →
86	Radeon RX 7900 XTX 24GB	2	85 %	~404 tok/s	36.5 tok/s	—	—	0.7 kW	Specs →
87	L4 24GB	2	85 %	~228 tok/s	11.4 tok/s	$1.18	$1.44	0.1 kW	Rent →
88	L20 48GB	1	83 %	~201 tok/s	17.3 tok/s	—	—	0.3 kW	Specs →
89	Radeon AI PRO R9600D 32GB	2	63 %	~304 tok/s	24.3 tok/s	—	—	0.3 kW	Specs →
90	Arc Pro B65 32GB	2	63 %	~297 tok/s	23.1 tok/s	—	—	0.4 kW	Specs →
91	Arc Pro B60 24GB	2	85 %	~262 tok/s	17.3 tok/s	—	—	0.4 kW	Specs →
92	L2 24GB	2	85 %	~208 tok/s	11.4 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB	2	85 %	~413 tok/s	16.4 tok/s	—	—	0.1 kW	Specs →
94	MLU370-X8	1	83 %	~154 tok/s	12.3 tok/s	—	—	0.3 kW	Specs →
95	Radeon PRO W7800 48GB	1	83 %	~167 tok/s	17.3 tok/s	—	—	0.3 kW	Specs →
96	Radeon PRO W7800 32GB	2	63 %	~276 tok/s	21.9 tok/s	—	—	0.5 kW	Specs →
97	GeForce RTX 3090 24GB	2	85 %	~272 tok/s	35.6 tok/s	$1.00	$1.02	0.7 kW	Rent →
98	A10G 24GB	2	85 %	~237 tok/s	22.8 tok/s	$2.01	$2.35	0.6 kW	Rent →
99	Atlas 300V Pro	1	83 %	~76.8 tok/s	4.1 tok/s	—	—	—	Specs →
100	Atlas 300I Pro	2	85 %	~146 tok/s	7.8 tok/s	—	—	—	Specs →
101	Atlas 300V	2	85 %	~125 tok/s	7.8 tok/s	—	—	0.1 kW	Specs →
102	GeForce RTX 3090 Ti 24GB	2	85 %	~173 tok/s	38.3 tok/s	$0.92	$1.48	0.9 kW	Rent →
103	Radeon PRO W6800 32GB	2	63 %	~139 tok/s	19.5 tok/s	—	—	0.5 kW	Specs →
104	RTX A5000 24GB	2	85 %	~122 tok/s	29.2 tok/s	$0.54	$1.23	0.5 kW	Rent →
105	Radeon PRO V710 28GB	2	72 %	~111 tok/s	17.0 tok/s	—	—	0.3 kW	Specs →

By vendor

At INT4 / FP4 the NVIDIA pick is 1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 920 tokens/s for $0.50/hr; the AMD pick is 1 x Instinct MI355X OAM at about 4.3k tokens/s for $5.99/hr. Per rental dollar NVIDIA delivers 2.57x the tokens of AMD here.

NVIDIA 51 parts fit

Best value: 1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 920 tok/s for $0.50/hr

Fewest GPUs: 1 x Rubin SXM, 12k tok/s

AMD 18 parts fit

Best value: 1 x Instinct MI355X OAM, 4.3k tok/s for $5.99/hr

Fewest GPUs: 1 x Instinct MI455X, 13k tok/s

Huawei 10 parts fit

Fewest GPUs: 1 x Ascend 950DT, 1.7k tok/s

Intel 7 parts fit

Fewest GPUs: 1 x Gaudi 3 128GB, 1.5k tok/s

Google 4 parts fit

Fewest GPUs: 1 x TPU v7 192GB, 2.7k tok/s

AWS 2 parts fit

Fewest GPUs: 2 x Inferentia2, 488 tok/s

Other 7 parts fit

Fewest GPUs: 1 x BR100, 743 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	1.41M tok/s	$2,480	700 kW	$0.49
B	100 x GB200 NVL72 GPU 186GB	100 x TP1	0	429k tok/s	$1,050	—	$0.68

Size gpt-oss-20b 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 gpt-oss-20b need?

The weights take 11.8 GB at the native INT4 / FP4 precision (41.8 GB at BF16, 20.9 GB at FP8, 11.8 GB at INT4). Each concurrent stream adds 48 KB of KV cache per token: 0.1 GB at 4,096 tokens and 0.8 GB at 32,768 tokens.

What is the bare minimum to run gpt-oss-20b?

1 x GB300 NVL72 GPU 288GB at INT4 / FP4 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 1 x RTX A5000 24GB at $0.27 per hour.

How many H100s do you need to run gpt-oss-20b?

one H100 at INT4 / FP4 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 1 H100s across 1 replicas.

How much does gpt-oss-20b cost per million tokens on an H100?

About $0.43 per million output tokens at INT4 / FP4, 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 $0.15.

What precision does gpt-oss-20b ship in?

The published checkpoint is MXFP4. That 4-bit format is what the lab validated, so the INT4 column is the native one; BF16 figures describe a dequantised copy nobody would deploy.

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. Active parameters per the OpenAI model card. Half the layers attend over a 128-token sliding window.