Microsoft Phi-4

Phi-4 hardware requirements

Fits on one Instinct MI355X OAM at BF16. For 1,000 tokens/s at 16k context you need 4 x H100 SXM5 80GB at $7.32/hr or 6 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $3.00/hr.

Parameters

14.7B

Active per token

All

dense model

KV cache per token

200 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Phi-4

The strongest part each vendor offers for Phi-4 at BF16 with 16,384 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 16,384 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 2 x Inferentia2

Memory fills at: 1 x GB300 NVL72 GPU 288GB 73 users, 1 x Instinct MI355X OAM 73 users, 1 x Ascend 910C 27 users, 1 x Gaudi 2 96GB 18 users, 1 x TPU v7 192GB 46 users, 2 x Inferentia2 9 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

1 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 196
Per hour: $5.99
Per million tokens: $8.48
Memory used: 11%

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: 43.8
Per hour: $0.50
Per million tokens: $3.17
Memory used: 33%

Rent on RunPod

Cheapest per million tokens

2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 16,384 tokens, 32 streams

Tokens/s: 387
Per hour: $1.00
Per million tokens: $0.72
Memory used: 75%

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	33 %	2.5k tok/s	128 tok/s	—	—	—	Specs →
2	Rubin SXM	1	49 %	2.2k tok/s	121 tok/s	—	—	—	Specs →
3	TPU 8t announced	1	66 %	~899 tok/s	35.8 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	49 %	~1.0k tok/s	47.2 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	49 %	987 tok/s	43.9 tok/s	$5.99	$1.69	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	49 %	985 tok/s	43.9 tok/s	$12.42	$3.50	—	Rent →
7	GB200 NVL72 GPU 186GB	1	77 %	985 tok/s	43.9 tok/s	$10.50	$2.96	—	Rent →
8	Instinct MI350X OAM	1	49 %	962 tok/s	43.9 tok/s	$6.16	$1.78	1.0 kW	Rent →
9	TPU v7 192GB	1	74 %	909 tok/s	40.5 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	33 %	~1.7k tok/s	128 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	54 %	954 tok/s	43.9 tok/s	$4.99	$1.45	1.4 kW	Rent →
12	B200 SXM 180GB	1	79 %	929 tok/s	42.2 tok/s	$4.09	$1.22	1.0 kW	Rent →
13	Ascend 970 announced	1	49 %	1.3k tok/s	79.0 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	1	74 %	874 tok/s	43.9 tok/s	—	—	0.7 kW	Specs →
15	Jalapeño announced	1	66 %	~1.2k tok/s	84.5 tok/s	—	—	0.7 kW	Specs →
16	Instinct MI325X OAM	1	55 %	654 tok/s	32.9 tok/s	$3.80	$1.61	1.0 kW	Rent →
17	Instinct MI300X 192GB	1	74 %	605 tok/s	29.1 tok/s	$2.39	$1.10	0.8 kW	Rent →
18	BR100	4	56 %	837 tok/s	31.6 tok/s	—	—	2.2 kW	Specs →
19	Ascend 960 announced	1	49 %	740 tok/s	52.7 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	2	51 %	974 tok/s	50.0 tok/s	$4.96	$1.42	1.4 kW	Rent →
21	H100 SXM5 80GB	2	90 %	771 tok/s	34.9 tok/s	$3.66	$1.32	1.4 kW	Rent →
22	TPU v6e 32GB	8	57 %	1.6k tok/s	61.1 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	2	56 %	719 tok/s	34.2 tok/s	—	—	1.2 kW	Specs →
24	H200 NVL 141GB	2	51 %	~908 tok/s	50.0 tok/s	$7.23	$2.21	1.2 kW	Rent →
25	H100 NVL 94GB	2	76 %	~803 tok/s	40.7 tok/s	$6.38	$2.21	0.8 kW	Rent →
26	Ascend 910C	2	56 %	697 tok/s	33.4 tok/s	—	—	1.2 kW	Specs →
27	H100 PCIe 80GB	2	90 %	494 tok/s	20.8 tok/s	$5.00	$2.81	0.7 kW	Rent →
28	H800 PCIe 80GB	2	90 %	494 tok/s	20.8 tok/s	—	—	0.7 kW	Specs →
29	Ascend 950DT	1	99 %	356 tok/s	21.9 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	2	75 %	406 tok/s	18.7 tok/s	$3.60	$2.46	1.2 kW	Rent →
31	Ascend 950PR	2	56 %	374 tok/s	16.7 tok/s	—	—	1.8 kW	Specs →
32	MTT S5000	2	90 %	~374 tok/s	16.7 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	2	75 %	374 tok/s	16.6 tok/s	$1.38	$1.03	1.2 kW	Rent →
34	Gaudi 3 128GB	2	56 %	954 tok/s	38.6 tok/s	—	—	1.8 kW	Specs →
35	TPU v5p 95GB	2	75 %	512 tok/s	28.8 tok/s	—	—	0.9 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	2	75 %	387 tok/s	18.7 tok/s	$1.00	$0.72	0.6 kW	Rent →
37	Cloud AI 100 Ultra	2	56 %	~157 tok/s	5.7 tok/s	—	—	0.3 kW	Specs →
38	Gaudi 2 96GB	2	75 %	466 tok/s	25.5 tok/s	$2.20	$1.31	1.2 kW	Rent →
39	Atlas 350	2	64 %	325 tok/s	14.6 tok/s	—	—	1.2 kW	Specs →
40	Data Center GPU Max 1100 48GB	4	75 %	559 tok/s	24.3 tok/s	—	—	1.2 kW	Specs →
41	GeForce RTX 5090 32GB	8	57 %	1.0k tok/s	66.9 tok/s	$9.52	$2.64	4.6 kW	Rent →
42	Ascend 910B	4	56 %	660 tok/s	31.6 tok/s	—	—	1.6 kW	Specs →
43	MI250X 128GB	2	56 %	511 tok/s	34.2 tok/s	—	—	1.1 kW	Specs →
44	RTX 6000 Ada 48GB	4	75 %	~449 tok/s	19.0 tok/s	$3.88	$2.40	1.2 kW	Rent →
45	Instinct MI250 128GB	2	56 %	496 tok/s	34.2 tok/s	—	—	1.0 kW	Specs →
46	L40S 48GB	4	75 %	415 tok/s	17.1 tok/s	$3.04	$2.04	1.4 kW	Rent →
47	A800 PCIe 80GB	2	90 %	354 tok/s	20.2 tok/s	—	—	0.6 kW	Specs →
48	A800 SXM4	2	90 %	364 tok/s	21.3 tok/s	—	—	0.8 kW	Specs →
49	A100 PCIe 80GB	2	90 %	354 tok/s	20.2 tok/s	$2.70	$2.12	0.6 kW	Rent →
50	A100 SXM4 80GB	2	90 %	364 tok/s	21.3 tok/s	$2.80	$2.14	0.8 kW	Rent →
51	A100 PCIe 40GB	4	91 %	591 tok/s	30.7 tok/s	—	—	1.0 kW	Specs →
52	A100 SXM4 40GB	4	91 %	591 tok/s	30.7 tok/s	$6.24	$2.93	1.6 kW	Rent →
53	GeForce RTX 5090 D V2 24GB	8	77 %	~1.0k tok/s	50.1 tok/s	—	—	4.6 kW	Specs →
54	RTX 5880 Ada 48GB	4	75 %	~415 tok/s	19.0 tok/s	—	—	1.1 kW	Specs →
55	TPU v4 32GB	8	57 %	907 tok/s	44.8 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB	8	57 %	~532 tok/s	21.5 tok/s	$6.64	$3.47	2.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	4	75 %	~502 tok/s	26.5 tok/s	$3.84	$2.13	1.2 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	2	100 %	~265 tok/s	14.0 tok/s	—	—	0.6 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition	8	57 %	627 tok/s	29.8 tok/s	—	—	1.3 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB	8	57 %	674 tok/s	33.4 tok/s	$5.76	$2.37	1.6 kW	Rent →
61	Radeon AI PRO R9700 32GB	8	57 %	529 tok/s	23.9 tok/s	—	—	2.4 kW	Specs →
62	Inferentia2	8	57 %	622 tok/s	30.6 tok/s	—	—	—	Specs →
63	Trainium	8	57 %	622 tok/s	30.6 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB	8	57 %	530 tok/s	45.8 tok/s	—	—	2.4 kW	Specs →
65	Arc Pro B70 32GB	8	57 %	~504 tok/s	22.7 tok/s	—	—	1.8 kW	Specs →
66	L40 48GB	4	75 %	334 tok/s	17.1 tok/s	$3.28	$2.73	1.2 kW	Rent →
67	Instinct MI210 PCIe	4	56 %	470 tok/s	32.4 tok/s	—	—	1.2 kW	Specs →
68	GeForce RTX 4090 24GB	8	77 %	664 tok/s	37.6 tok/s	$7.12	$2.98	3.6 kW	Rent →
69	A30 24GB	8	77 %	636 tok/s	34.8 tok/s	—	—	1.3 kW	Specs →
70	RTX 4500 Ada 24GB	8	77 %	~379 tok/s	16.1 tok/s	—	—	1.7 kW	Specs →
71	RTX A6000 48GB	4	75 %	292 tok/s	15.2 tok/s	$2.00	$1.90	1.2 kW	Rent →
72	A40 48GB	4	75 %	272 tok/s	13.7 tok/s	$2.36	$2.41	1.2 kW	Rent →
73	H20 141GB HBM3e	2	51 %	290 tok/s	41.7 tok/s	—	—	0.8 kW	Specs →
74	H20 96GB	2	75 %	290 tok/s	41.7 tok/s	—	—	0.8 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB	8	77 %	~499 tok/s	25.1 tok/s	$4.56	$2.54	1.2 kW	Rent →
76	BI-V100 Tiangai 100	8	57 %	687 tok/s	44.8 tok/s	—	—	2.0 kW	Specs →
77	Atlas 300I Duo 48GB	4	75 %	186 tok/s	8.1 tok/s	—	—	0.6 kW	Specs →
78	Atlas 300I Duo 96GB	2	75 %	98.1 tok/s	4.3 tok/s	—	—	0.3 kW	Specs →
79	RTX A5500 24GB	8	77 %	~524 tok/s	28.7 tok/s	—	—	1.8 kW	Specs →
80	Wormhole n300s	8	77 %	434 tok/s	21.5 tok/s	—	—	2.4 kW	Specs →
81	V100S PCIe 32GB	8	57 %	627 tok/s	42.3 tok/s	$7.04	$3.12	2.0 kW	Rent →
82	ATOM-Max	4	56 %	314 tok/s	20.2 tok/s	—	—	1.4 kW	Specs →
83	A10 24GB	8	77 %	437 tok/s	22.4 tok/s	$8.81	$5.60	1.2 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	4	75 %	282 tok/s	17.1 tok/s	—	—	1.2 kW	Specs →
85	Radeon PRO W7900 48GB	4	75 %	282 tok/s	17.1 tok/s	—	—	1.2 kW	Specs →
86	Radeon RX 7900 XTX 24GB	8	77 %	562 tok/s	35.8 tok/s	—	—	2.8 kW	Specs →
87	L4 24GB	8	77 %	269 tok/s	11.2 tok/s	$4.72	$4.87	0.6 kW	Rent →
88	L20 48GB	4	75 %	278 tok/s	17.1 tok/s	—	—	1.1 kW	Specs →
89	Radeon AI PRO R9600D 32GB	8	57 %	411 tok/s	23.9 tok/s	—	—	1.2 kW	Specs →
90	Arc Pro B65 32GB	8	57 %	~399 tok/s	22.7 tok/s	—	—	1.6 kW	Specs →
91	Arc Pro B60 24GB	8	77 %	~337 tok/s	17.0 tok/s	—	—	1.6 kW	Specs →
92	L2 24GB	8	77 %	253 tok/s	11.2 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB	8	77 %	~323 tok/s	16.1 tok/s	—	—	0.6 kW	Specs →
94	MLU370-X8	4	75 %	210 tok/s	12.1 tok/s	—	—	1.0 kW	Specs →
95	Radeon PRO W7800 48GB	4	75 %	240 tok/s	17.1 tok/s	—	—	1.0 kW	Specs →
96	Radeon PRO W7800 32GB	8	57 %	372 tok/s	21.5 tok/s	—	—	2.1 kW	Specs →
97	GeForce RTX 3090 24GB	8	77 %	406 tok/s	34.9 tok/s	$4.00	$2.74	2.8 kW	Rent →
98	A10G 24GB	8	77 %	335 tok/s	22.4 tok/s	$8.05	$6.68	2.4 kW	Rent →
99	Atlas 300V Pro	4	75 %	93.2 tok/s	4.0 tok/s	—	—	—	Specs →
100	Atlas 300I Pro	8	77 %	176 tok/s	7.6 tok/s	—	—	—	Specs →
101	Atlas 300V	8	77 %	158 tok/s	7.6 tok/s	—	—	0.6 kW	Specs →
102	GeForce RTX 3090 Ti 24GB	8	77 %	277 tok/s	37.6 tok/s	$3.68	$3.70	3.6 kW	Rent →
103	Radeon PRO W6800 32GB	8	57 %	210 tok/s	19.1 tok/s	—	—	2.0 kW	Specs →
104	RTX A5000 24GB	8	77 %	196 tok/s	28.7 tok/s	$2.16	$3.06	1.8 kW	Rent →
105	Radeon PRO V710 28GB	8	66 %	170 tok/s	16.7 tok/s	—	—	1.3 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 387 tokens/s for $1.00/hr; the AMD pick is 1 x Instinct MI300X 192GB at about 605 tokens/s for $2.39/hr. Per rental dollar NVIDIA delivers 1.53x the tokens of AMD here.

NVIDIA 51 parts fit

Best value: 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 387 tok/s for $1.00/hr

Fewest GPUs: 1 x Rubin SXM, 2.2k tok/s

AMD 18 parts fit

Best value: 1 x Instinct MI300X 192GB, 605 tok/s for $2.39/hr

Fewest GPUs: 1 x Instinct MI455X, 2.5k tok/s

Huawei 10 parts fit

Fewest GPUs: 1 x Ascend 950DT, 356 tok/s

Intel 7 parts fit

Best value: 2 x Gaudi 2 96GB, 466 tok/s for $2.20/hr

Fewest GPUs: 2 x Gaudi 3 128GB, 954 tok/s

Google 4 parts fit

Fewest GPUs: 1 x TPU v7 192GB, 909 tok/s

AWS 2 parts fit

Fewest GPUs: 8 x Inferentia2, 622 tok/s

Other 7 parts fit

Fewest GPUs: 2 x MTT S5000, 374 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	487k tok/s	$2,480	700 kW	$1.42
B	100 x GB200 NVL72 GPU 186GB	100 x TP1	0	99k tok/s	$1,050	—	$2.96

Size Phi-4 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 Phi-4 need?

The weights take 29.3 GB at the native BF16 precision (29.3 GB at BF16, 14.7 GB at FP8, 8.2 GB at INT4). Each concurrent stream adds 200 KB of KV cache per token: 0.8 GB at 4,096 tokens and 3.4 GB at 16,384 tokens.

What is the bare minimum to run Phi-4?

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 Phi-4?

one H100 at BF16 for 4,096 tokens of context and one stream. For 16,384 tokens and 32 concurrent streams the tensor-parallel group is 2 H100s, and 1,000 tokens per second takes 4 H100s across 2 replicas.

How much does Phi-4 cost per million tokens on an H100?

About $1.32 per million output tokens at BF16, 16,384 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 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $0.72.

What precision does Phi-4 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.