GPU rental for neural networks
Training a neural network is rarely blocked by the idea or the code — it is blocked by access to a card of the right class. Buying your own A100 or H100 server pays off only under constant load, while experiments come in waves: a week of runs, then a pause for data labelling and architecture fixes. QuData closes that gap — you take a card for exactly as long as it is actually computing and release it when training is done.
The marketplace collects offers from verified providers: from A10 and RTX 4090 for pipeline debugging to H100 and H200 for heavy runs. Billing is per minute, and an environment with CUDA, PyTorch and TensorFlow spins up from a ready-made image in about 30 seconds. No hardware procurement, no delivery queue, no idle cards between experiments.
Ready-made images
Setting up drivers, CUDA versions and compatible framework builds eats hours before the first line of training. We maintain ready-made environments — pick one when launching the instance and connect over SSH right away.
PyTorch + CUDA
A current PyTorch build with a matching CUDA and cuDNN version. The default image for most training workloads.TensorFlow
An environment with GPU acceleration and Keras. Suits projects that inherited a TF codebase.JupyterLab
Interactive notebooks straight in the browser. Convenient for the research phase and hyperparameter tuning.Docker-ready
A clean instance with NVIDIA Container Toolkit configured, if you bring your own image.Who it suits
ML engineers
Run several training configurations in parallel on different cards and keep only the winner. Renting lets you scale an experiment horizontally without begging for a hardware budget.Startups without their own hardware
Validate a hypothesis and reach a working model without freezing money in equipment before the product has proven demand.Universities and labs
Close a project task or a course without capital expenditure or procurement approval, staying within the grant or semester deadline.Product teams
Fine-tune a model for your own domain on your own data without sending it to external APIs.How to choose a GPU for neural network training
Choosing a card for training comes down to three parameters: video memory size, mixed-precision compute speed and memory bandwidth. The first one is decisive: if the model together with the optimizer and the batch does not fit into VRAM, training simply will not start, however fast the chip is.
How much video memory neural network training needs
A practical rule of thumb: training needs three to four times more memory than the model weights themselves occupy. Besides the weights, memory holds gradients, optimizer states and activations of intermediate layers. So a model that fits into 12 GB for inference can easily demand 40 GB or more during training.
For convolutional networks, classic computer vision tasks and small transformers, 24 GB cards are enough — RTX 4090 or A10. For mid-sized transformers and multimodal architectures aim for 48 GB: L40S or A40. Large language and generative models require 80 GB and above, meaning A100, H100 or H200.
Technical tricks lower the requirements: mixed-precision training, gradient checkpointing, gradient accumulation instead of a larger batch. They save memory at the cost of speed, and on a rented card that trade-off is measured directly — sometimes a higher-class card that finishes faster is cheaper than saving on the hardware tier.
Debugging and a full run need different cards
A typical mistake is taking the top card from day one. At the stage where you are catching bugs in the data pipeline and checking that the metric is computed at all, raw power is not needed: an inexpensive card will do, and sometimes a trimmed dataset too. The full run is launched deliberately, on a card matched to the task, and it is a separate short session.
Per-minute billing makes this two-stage scheme pay off. You do not pay for idle time between runs and do not keep an expensive card while fixing code.
One card or several
A multi-GPU configuration is needed when the model does not fit into the memory of a single accelerator or when training time is critical. With data parallelism each card computes its slice of the batch, and time shrinks almost proportionally to the number of cards. With model parallelism the layers are distributed across cards — that mode is harder to set up but allows training what physically does not fit into one accelerator.
For most applied tasks a single card of the right class is enough, and moving to multi-GPU is justified only once you hit the ceiling.
What to consider before you start
Check framework and CUDA version compatibility — ready-made images remove that question. Plan in advance where checkpoints go: instance data is deleted after the session, so exporting intermediate results must be part of the script rather than a manual step at the end.
Frequently asked questions
Which GPU do I need to train a neural network?
It depends on the model size. For experiments and small networks an RTX 4090 or A10 is enough; for models with billions of parameters you need an A100 or H100 with 80 GB of video memory. The table above shows which workload fits on each card.
How much does it cost to train a neural network on a rented GPU?
Billing is per minute, so the final cost equals training time multiplied by the current hourly price of the chosen card. Fine-tuning a small model on an A100 takes a few hours; full training from scratch is estimated individually — see the marketplace for current rates.
Do I have to configure CUDA and drivers myself?
No. Instances launch from ready-made images with CUDA, PyTorch and TensorFlow preinstalled — the environment is ready to work right after start.
Can I rent a GPU by the hour instead of by the month?
Yes, the minimum rental period is one hour and billing is per minute. This is the main scenario for experiments: you pay only for actual compute time.
What happens to my data after the rental ends?
All data is deleted once the session ends, and while the instance runs it is encrypted — the host has no access to it either during or after the rental.