Rent the GPU for your next experiment.

Capacity for your campaign, without buying a card.

Mobilize your project's GPUs for 3, 7 or 30 days. You choose a configuration, a period and a number of lots: each listing brings together the memory per card, the available stock and the package price so you can prepare your purchase.

One lot contains one card for fourteen configurations and two cards for B200. This unit lets you compare an experiment on one GPU and a campaign spread across several lots, with a clearly defined total.

Find the cards worth choosing for your model.

An inference validation, a model adaptation and a series of training runs do not require the same memory. Start from your workload and your libraries, then compare the packages of the cards that meet those needs. IteraGPU gives you several options at every stage of the campaign.

• Start with 24 GB per card: the NVIDIA RTX A5000 24GB gives access to the entry price. The RTX 3090, RTX 4090 and L4 also offer 24 GB of capacity, with their own packages.
• Move up to 48 GB on a single card: compare A40, RTX A6000, RTX 6000 Ada and L40S to accommodate a workload that exceeds the capacity of a 24 GB card.
• Plan for at least 80 GB per card: look at A100 SXM 80 GB, H100 PCIe or SXM 80 GB, then H200 SXM 141 GB depending on your environment and the memory required.
• Choose a large capacity or several GPUs: MI300X offers 192 GB on an AMD card; one B200 lot brings together two NVIDIA cards of 180 GB each. Go with the organization your software can make use of.

A duration for every stage of your research.

Three days for a targeted validation; seven days to compare a series of variants; thirty days to organize a campaign with several stages. Choose the window that matches your work plan and see the package total before ordering.

For independent experiments, you can plan a single lot used in succession or several lots for parallel runs. Two standard lots represent two GPUs; two B200 lots represent four. The two B200 cards are already included in the price of each lot.

Stay in control of your research, specify your needs.

You choose your models, your code and your dependencies. To select an offer, rely on the specifications of the listing and the requirements of your campaign.

If your decision depends on Jupyter access, file persistence, a preinstalled software version, RAM, storage, GPU interconnect, or provisioning time, have these points confirmed before ordering. Memory remains specific to each card: your software must organize the work across multiple GPUs.

Plans buy a period of capacity; throughput and the number of completed experiments depend on your workload. If your project requires hourly billing or a performance commitment, specify this condition before choosing.