Rent a cloud GPU by the hour

55 GPU types are on the board right now, from USD 0.09 an hour. Every price is today’s lowest on a verified host, billed by the second. Pick a card to see its memory, what it is good for, what common jobs cost on it, and its price over the last weeks.

Hopper — datacenter

NVIDIA's datacenter flagship behind most large-model training today, with very high memory bandwidth and FP8 support.

Ampere — datacenter

The A100 generation: still a workhorse for training and serving, with fast HBM memory.

Ada — datacenter

Datacenter cards from the Ada generation, strong at inference and at graphics-heavy AI work.

Ada — workstation

Professional workstation cards with large memory, suited to long runs and big models.

Blackwell — consumer

The newest GeForce generation: very fast for the price, with FP8 and FP4 support.

Ada — consumer

GeForce RTX 40 series: excellent value for fine-tuning, image generation and inference.

Ampere — workstation

Professional RTX A-series cards with generous memory for the price.

Ampere — consumer

GeForce RTX 30 series: the budget choice for learning, inference and image work.

Turing

Older RTX and T4-class cards: fine for learning, notebooks and light inference.

Volta

The V100 generation: older, but its fast HBM memory still makes it useful for training.

Pascal

Older GTX-era cards: the cheapest way to learn, though without tensor cores.

Maxwell

Very old cards: the cheapest of all, for experiments where speed does not matter.

Mining cards

Cards originally built for crypto mining, with limited PCIe bandwidth.

Other

A GPU available on the live board.