How to use PetaPrice

Find the cheapest GPU for your workload in a few steps, and what to look for when choosing one.

Using PetaPrice step by step

Choose what to optimize for
On the home page, pick FLOPS for training and compute-bound work, VRAM to fit the biggest model per dollar, or Bandwidth for LLM token generation. For FLOPS, also pick the precision you run in, usually FP16/BF16.
Filter to GPUs that fit
Set a minimum VRAM so your model fits, the number of GPUs per instance, and on-demand, spot or reserved pricing. You can also filter by vendor, provider headquarters (EU only) and managed cloud vs. community hosts.
Read the ranking
The ranking sorts GPUs by value for money at their cheapest matching offer. The label shows how much more each GPU costs per unit than the best. The sweet-spot chart plots price against performance: up and to the left is better.
Open a GPU
A GPU's page lists every provider renting it, each configuration with its price, a daily price history and the cost at every precision. Tap a provider to see its configurations and the link to rent.
Compare GPUs side by side
Pick up to four GPUs in Compare to see prices, compute per dollar, memory, bandwidth and providers next to each other.
Check the price and rent
Prices are indicative list prices refreshed about every 2 hours. Follow the Rent link and confirm the price on the provider's website before you book.

What matters for each kind of job, and the GPUs that are the best value for it right now.

What matters: Enough memory for weights and KV cache, then memory bandwidth: generating tokens is bandwidth-bound.

Best value right now · Best bandwidth per dollar with 80 GB or more

What matters: Dense BF16 or FP8 compute, memory for weights, gradients and optimizer state, and NVLink between GPUs.

Best value right now · Best FP8 compute per dollar, datacenter GPUs

What matters: A low hourly price and at least 24 GB. Consumer and workstation cards are often the cheapest way to start.

Best value right now · Cheapest GPU-hour with 24 GB or more

What matters: Double-precision throughput. Many AI GPUs, including Blackwell and all consumer cards, have little of it.

Best value right now · Best FP64 compute per dollar

Lower precision is faster and needs less memory, if your GPU runs it natively and your model tolerates it.

Format	Bits	Used for	Native on
FP64	64	Scientific simulation	Any
FP32	32	Non-tensor work, numerically sensitive code	Any
TF32	19	FP32 training on tensor cores	Blackwell Ultra, Blackwell, Hopper, Ada Lovelace, CDNA 3, Xe-HPC, Ampere
FP16 / BF16	16	The standard for training and serving LLMs	Blackwell Ultra, Blackwell, CDNA 4, Hopper, Ada Lovelace, CDNA 3, Gaudi, TPU, Trainium, RDNA 3, Xe-HPC, Inferentia, Ampere, CDNA 2, CDNA
FP8	8	Faster training and inference on recent GPUs	Blackwell Ultra, Blackwell, CDNA 4, Hopper, Ada Lovelace, CDNA 3, Gaudi, Trainium, Inferentia
INT8	8	Quantized inference	Blackwell Ultra, Blackwell, CDNA 4, Hopper, Ada Lovelace, CDNA 3, TPU, RDNA 3, Xe-HPC, Inferentia, Ampere, CDNA 2, CDNA, Turing
FP4	4	Quantized inference of large models	Blackwell Ultra, Blackwell, CDNA 4

Filters are saved in the URL, so you can bookmark or share any view.
Spot capacity is often half the price or less, but can be stopped at any time. Use it for jobs that checkpoint.
Compare dense FLOPS at the precision you actually use: FP8 and FP4 numbers only help if your software uses them.
Each provider's page shows every GPU it rents and how its prices compare with the cheapest provider.