Methodology

The numbers

$/PFLOP·h = price per GPU-hour ÷ peak PFLOPS at the precision you choose. For example, an H100 SXM at $2.00/h delivers 0.989 dense BF16 PFLOPS, so it costs $2.02 per PFLOP·h. All compute figures on PetaPrice are in PFLOPS.
Always dense. Vendors often quote 2:4-sparsity numbers, which double the headline figure. LLM training and inference almost never use structured sparsity, so we halve those numbers. Dense and sparse figures are never mixed.
GeForce cards use the FP16/BF16 tensor rate with FP32 accumulate, which is what mixed-precision training actually runs at. NVIDIA markets GeForce with the FP16-accumulate rate, which is twice as high.
These are datasheet peaks. Real model FLOPS utilization is usually 30–60%. Use the numbers to compare GPUs with each other, not to predict wall-clock time.
Two alternative lenses: $/GB·h of VRAM (what's the cheapest way to fit a model?) and $/TB/s·h of memory bandwidth (LLM decoding speed is usually bandwidth-bound).

The prices

Prices come directly from each provider wherever they publish them: public APIs (Vast.ai, RunPod, Verda, Vultr, Cudo, Hyperbolic, Scaleway, OVHcloud), official price lists (AWS, Azure, Oracle Cloud) or their own pricing pages (Lambda, Nebius, Hyperstack, DigitalOcean, Together AI, Crusoe, Hot Aisle, TensorWave, Cirrascale, Seeweb). Shadeform's live multi-cloud API and SkyPilot's public catalogs only fill in providers without a direct source, such as Google Cloud and IBM.
The data refreshes every 2 hours. Every price is normalized to USD per GPU per hour. The table also shows the price of the whole instance. Euro prices are converted at the European Central Bank's daily reference rate.
On-demand, spot (interruptible) and reserved/committed prices are never mixed. The default view is on-demand only.
Hyperscaler SKUs list the cheapest region, with a count of other regions. For marketplaces such as Vast.ai we fold identical listings together and show the cheapest three per configuration.
GPUs that providers list but that aren't in our spec sheet yet are shown as "specs pending" with their price and memory, so new hardware appears as soon as it's offered.
Prices exclude storage, egress and IP addresses unless the provider bundles them. Vast.ai prices include the host's default disk.

Spec sheet

Peak dense throughput per GPU. INT8 is in TOPS. A dash means the GPU has no native support for that precision. (PFLOPS)

GPU	Vendor	Architecture	VRAM	GB/s	FP16 / BF16	FP8	FP4	INT8	TF32	FP32	FP64
GB300	NVIDIA	Blackwell Ultra	288 GB	8,000	2.5	5	15	0.33	1.25	0.08	0.0013
B300 SXM	NVIDIA	Blackwell Ultra	288 GB	8,000	2.25	4.5	13.5	0.33	1.1	0.075	0.0013
GB200	NVIDIA	Blackwell	186 GB	8,000	2.5	5	10	5	1.25	0.08	0.045
B200 SXM	NVIDIA	Blackwell	180 GB	7,700	2.25	4.5	9	4.5	1.1	0.075	0.037
H200 SXM	NVIDIA	Hopper	141 GB	4,800	0.989	1.98	—	1.98	0.495	0.067	0.067
H200 NVL	NVIDIA	Hopper	141 GB	4,800	0.835	1.67	—	1.67	0.417	0.06	0.06
GH200	NVIDIA	Hopper	96 GB	4,000	0.989	1.98	—	1.98	0.495	0.067	0.067
H100 SXM	NVIDIA	Hopper	80 GB	3,350	0.989	1.98	—	1.98	0.495	0.067	0.067
H100 NVL	NVIDIA	Hopper	94 GB	3,900	0.835	1.67	—	1.67	0.417	0.06	0.06
H100 PCIe	NVIDIA	Hopper	80 GB	2,000	0.756	1.51	—	1.51	0.378	0.051	0.051
H800 SXM	NVIDIA	Hopper	80 GB	3,350	0.989	1.98	—	1.98	0.495	0.067	0.001
H20	NVIDIA	Hopper	96 GB	4,000	0.148	0.296	—	0.296	0.074	0.044	0.001
A100 SXM 80GB	NVIDIA	Ampere	80 GB	2,039	0.312	—	—	0.624	0.156	0.0195	0.0195
A100 SXM 40GB	NVIDIA	Ampere	40 GB	1,555	0.312	—	—	0.624	0.156	0.0195	0.0195
A100 PCIe 80GB	NVIDIA	Ampere	80 GB	1,935	0.312	—	—	0.624	0.156	0.0195	0.0195
A100 PCIe 40GB	NVIDIA	Ampere	40 GB	1,555	0.312	—	—	0.624	0.156	0.0195	0.0195
A800	NVIDIA	Ampere	80 GB	1,935	0.312	—	—	0.624	0.156	0.0195	0.0195
L40S	NVIDIA	Ada Lovelace	48 GB	864	0.362	0.733	—	0.733	0.183	0.0916	0.0014
L40	NVIDIA	Ada Lovelace	48 GB	864	0.181	0.362	—	0.362	0.0905	0.0905	0.0014
L4	NVIDIA	Ada Lovelace	24 GB	300	0.121	0.242	—	0.242	0.06	0.0303	0.0005
A40	NVIDIA	Ampere	48 GB	696	0.15	—	—	0.299	0.0748	0.0374	0.0006
A30	NVIDIA	Ampere	24 GB	933	0.165	—	—	0.33	0.082	0.0103	0.0103
A10	NVIDIA	Ampere	24 GB	600	0.125	—	—	0.25	0.0625	0.0312	0.0005
A10G	NVIDIA	Ampere	24 GB	600	0.07	—	—	0.14	0.035	0.0312	0.0005
A16	NVIDIA	Ampere	16 GB	200	0.018	—	—	0.036	0.009	0.0045	—
T4	NVIDIA	Turing	16 GB	320	0.065	—	—	0.13	—	0.0081	0.0003
V100 SXM2 32GB	NVIDIA	Volta	32 GB	900	0.125	—	—	—	—	0.0157	0.0078
V100 SXM2 16GB	NVIDIA	Volta	16 GB	900	0.125	—	—	—	—	0.0157	0.0078
V100 PCIe	NVIDIA	Volta	16 GB	900	0.112	—	—	—	—	0.014	0.007
V100S PCIe	NVIDIA	Volta	32 GB	1,134	0.13	—	—	—	—	0.0164	0.0082
P100	NVIDIA	Pascal	16 GB	732	0.0187	—	—	—	—	0.0093	0.0047
RTX PRO 6000 Blackwell	NVIDIA	Blackwell	96 GB	1,792	0.504	1.01	2.02	1.01	0.126	0.126	0.002
RTX 6000 Ada	NVIDIA	Ada Lovelace	48 GB	960	0.364	0.729	—	0.729	0.182	0.0911	0.0014
RTX 5000 Ada	NVIDIA	Ada Lovelace	32 GB	576	0.261	0.522	—	0.522	0.131	0.0653	0.001
RTX 4000 Ada	NVIDIA	Ada Lovelace	20 GB	360	0.0818	0.164	—	0.164	0.0409	0.0267	0.0004
RTX A6000	NVIDIA	Ampere	48 GB	768	0.155	—	—	0.31	0.0774	0.0387	0.0006
RTX A5000	NVIDIA	Ampere	24 GB	768	0.111	—	—	0.222	0.0556	0.0278	0.0004
RTX A4500	NVIDIA	Ampere	20 GB	640	0.0946	—	—	0.189	0.0473	0.0237	0.0004
RTX A4000	NVIDIA	Ampere	16 GB	448	0.0767	—	—	0.153	0.0384	0.0192	0.0003
RTX PRO 5000 Blackwell	NVIDIA	Blackwell	48 GB	1,344	0.268	0.535	1.07	0.535	0.067	0.0737	0.0012
RTX PRO 4000 Blackwell	NVIDIA	Blackwell	24 GB	672	0.147	0.294	0.589	0.294	0.037	0.0469	0.0007
RTX 5880 Ada	NVIDIA	Ada Lovelace	48 GB	960	0.277	0.555	—	0.555	0.139	0.0693	0.0011
RTX PRO 4500 Blackwell	NVIDIA	Blackwell	32 GB	896	0.211	0.422	0.843	0.422	0.053	0.0549	0.0009
RTX 4000 SFF Ada	NVIDIA	Ada Lovelace	20 GB	280	0.0767	0.153	—	0.153	0.0384	0.0192	0.0003
Quadro RTX 6000	NVIDIA	Turing	24 GB	672	0.131	—	—	0.261	—	0.0163	0.0005
RTX 2000 Ada	NVIDIA	Ada Lovelace	16 GB	224	0.048	0.096	—	0.096	0.024	0.012	0.0002
RTX A2000	NVIDIA	Ampere	12 GB	288	0.032	—	—	0.064	0.016	0.008	0.0001
Quadro RTX 8000	NVIDIA	Turing	48 GB	672	0.131	—	—	0.261	—	0.0163	0.0005
RTX 5090	NVIDIA	Blackwell	32 GB	1,792	0.209	0.419	1.68	0.838	0.105	0.105	0.0016
RTX 5080	NVIDIA	Blackwell	16 GB	960	0.113	0.225	0.9	0.45	0.0563	0.0563	0.0009
RTX 5070 Ti	NVIDIA	Blackwell	16 GB	896	0.0879	0.176	0.703	0.352	0.0439	0.0439	0.0007
RTX 4090	NVIDIA	Ada Lovelace	24 GB	1,008	0.165	0.33	—	0.661	0.0826	0.0826	0.0013
RTX 4080 SUPER	NVIDIA	Ada Lovelace	16 GB	736	0.104	0.209	—	0.418	0.0522	0.0522	0.0008
RTX 4080	NVIDIA	Ada Lovelace	16 GB	717	0.0975	0.195	—	0.39	0.0487	0.0487	0.0008
RTX 4070 Ti	NVIDIA	Ada Lovelace	12 GB	504	0.0802	0.16	—	0.321	0.0401	0.0401	0.0006
RTX 4070	NVIDIA	Ada Lovelace	12 GB	504	0.0583	0.117	—	0.233	0.0291	0.0291	0.0005
RTX 4060 Ti	NVIDIA	Ada Lovelace	16 GB	288	0.0441	0.0883	—	0.177	0.0221	0.0221	0.0003
RTX 3090 Ti	NVIDIA	Ampere	24 GB	1,008	0.08	—	—	0.32	0.04	0.04	0.0006
RTX 3090	NVIDIA	Ampere	24 GB	936	0.071	—	—	0.284	0.0356	0.0356	0.0006
RTX 3080 Ti	NVIDIA	Ampere	12 GB	912	0.0682	—	—	0.273	0.0341	0.0341	0.0005
RTX 3080	NVIDIA	Ampere	10 GB	760	0.0595	—	—	0.238	0.0298	0.0298	0.0005
RTX 3070	NVIDIA	Ampere	8 GB	448	0.0406	—	—	0.163	0.0203	0.0203	0.0003
RTX 3060	NVIDIA	Ampere	12 GB	360	0.0254	—	—	0.102	0.0127	0.0127	0.0002
RTX 2080 Ti	NVIDIA	Turing	11 GB	616	0.0538	—	—	0.215	—	0.0134	0.0004
RTX 4090D	NVIDIA	Ada Lovelace	24 GB	1,008	0.147	0.294	—	0.588	0.0735	0.0735	0.0011
RTX 5060 Ti	NVIDIA	Blackwell	16 GB	448	0.0474	0.0949	0.38	0.19	0.0237	0.0237	0.0004
RTX 4070 SUPER	NVIDIA	Ada Lovelace	12 GB	504	0.071	0.142	—	0.284	0.0355	0.0355	0.0006
RTX 3070 Ti	NVIDIA	Ampere	8 GB	608	0.0435	—	—	0.174	0.0217	0.0217	0.0003
RTX 3060 Ti	NVIDIA	Ampere	8 GB	448	0.0324	—	—	0.13	0.0162	0.0162	0.0003
Titan RTX	NVIDIA	Turing	24 GB	672	0.131	—	—	0.261	—	0.0163	0.0005
Instinct MI355X	AMD	CDNA 4	288 GB	8,000	2.52	5.03	10.1	5.03	—	0.157	0.0786
Instinct MI350X	AMD	CDNA 4	288 GB	8,000	2.31	4.61	9.23	4.61	—	0.144	0.0721
Instinct MI325X	AMD	CDNA 3	256 GB	6,000	1.31	2.61	—	2.61	0.654	0.163	0.163
Instinct MI300X	AMD	CDNA 3	192 GB	5,300	1.31	2.61	—	2.61	0.654	0.163	0.163
Instinct MI250X	AMD	CDNA 2	128 GB	3,277	0.383	—	—	0.383	—	0.0957	0.0957
Instinct MI250	AMD	CDNA 2	128 GB	3,277	0.362	—	—	0.362	—	0.0905	0.0905
Instinct MI210	AMD	CDNA 2	64 GB	1,638	0.181	—	—	0.181	—	0.0453	0.0453
Instinct MI100	AMD	CDNA	32 GB	1,229	0.185	—	—	0.185	—	0.0461	0.0115
Radeon PRO W7900	AMD	RDNA 3	48 GB	864	0.123	—	—	0.123	—	0.0613	0.0019
Radeon RX 7900 XTX	AMD	RDNA 3	24 GB	960	0.123	—	—	0.123	—	0.0614	0.0019
Gaudi 3	Intel	Gaudi	128 GB	3,700	1.83	1.83	—	—	—	0.0287	—
Gaudi 2	Intel	Gaudi	96 GB	2,450	0.432	0.865	—	—	—	0.011	—
Data Center GPU Max 1550	Intel	Xe-HPC	128 GB	3,277	0.839	—	—	1.68	0.419	0.0524	0.0524
TPU v6e (Trillium)	Google TPU	TPU	32 GB	1,640	0.918	—	—	1.84	—	—	—
TPU v5p	Google TPU	TPU	95 GB	2,765	0.459	—	—	0.918	—	—	—
TPU v5e	Google TPU	TPU	16 GB	819	0.197	—	—	0.394	—	—	—
Trainium2	AWS Trainium	Trainium	96 GB	2,900	0.667	1.3	—	—	—	0.181	—
Trainium	AWS Trainium	Trainium	32 GB	820	0.19	0.19	—	—	—	0.0475	—
Inferentia2	AWS Trainium	Inferentia	32 GB	820	0.19	0.19	—	0.38	—	0.0475	—