GPU rental for scientific computing and HPC
Scientific and engineering computations require capacity that is hard to justify in a purchase: supercomputer time is allocated by quota and comes with a queue, while an in-house cluster for a project task is economically unjustified. Yet the computation is needed here and now, not after a month of approvals.
QuData gives access to A100, H100 and H200 cards, including multi-GPU configurations, for the duration of a specific task. You get full root access over SSH and deploy the environment yourself — CUDA Toolkit, MPI, containers with computational packages. You pay for actual compute time, with no capital expenditure and no queue.
Ready-made images
In scientific computing the environment is usually specific to the task, so we provide the base and full control over the machine.
CUDA Toolkit
A clean instance with a current toolkit version, compilers and libraries. The starting point for your own code.MPI environment
A configured stack for parallel computations on multi-GPU configurations.Docker and Singularity
Running containers with computational packages, including images built inside your organization.JupyterLab
Interactive mode for data preparation, model debugging and result visualization.Who it suits
Universities and research institutes
Complete the computational part of a study within the grant deadline without going through procurement or waiting for a cluster quota.Engineering calculations
Run a series of CFD or structural analysis simulations for a specific project without maintaining capacity between jobs.Bioinformatics
Process sequencing data and run molecular modelling on capacity proportionate to the task.Financial modelling
Run Monte Carlo simulations and stress scenarios where computation time directly affects the value of the result.How to pick a GPU for scientific computing
The requirements of computational tasks differ fundamentally from those of machine learning — and a card that is optimal for training neural networks can be a poor choice for numerical modelling.
Double precision and why it decides everything
Most engineering and physical computations require double-precision arithmetic — FP64. Rounding error accumulates in iterative schemes, and single precision is often not enough for a stable solution.
This is where the main dividing line in the card line-up runs. The professional A100, H100 and H200 accelerators have full FP64 units, whereas gaming cards have double-precision performance cut by tens of times. The RTX 4090 is faster than the A100 on single-precision tasks and catastrophically slower in FP64. Before choosing a card, determine what precision your solver works in.
If the task allows mixed precision — and many modern schemes do — the picture changes and the choice becomes wider.
Memory and bandwidth
In numerical modelling, performance is more often limited by memory access speed than by compute cores. The computational mesh is re-read constantly, and memory bandwidth becomes decisive. Cards with HBM memory beat cards with GDDR precisely on such tasks, even at comparable peak performance.
Memory size determines the maximum problem size: the mesh must fit entirely into the card memory or be split correctly across several accelerators.
Multi-GPU and scaling
For tasks that do not fit on a single card, the computational domain is divided between accelerators. The efficiency of such scaling depends on how much data has to be transferred between cards on each iteration. Tasks with local interactions scale well; tasks with global couplings noticeably worse.
Before launching a long series of computations it makes sense to run a scaling test on two and four cards: it will show whether paying for a larger configuration is worthwhile or the gain will not justify the cost.
Organizing computation on rented capacity
Long computations should be split into stages with intermediate state saved. This is standard HPC practice, and on a rented instance it is especially appropriate: you can stop the machine between stages, restart the computation on a different configuration and not lose the result when the session ends. Plan the export of results to persistent storage as part of the computation script.
Frequently asked questions
What tasks can be computed on rented GPUs?
Molecular modelling, CFD and FEM engineering calculations, bioinformatics, Monte Carlo financial simulations and any task with massive parallelism.
Are multi-GPU configurations available?
Yes, the marketplace has multi-card builds with high-speed interconnect — they suit tasks that do not fit into the memory of a single accelerator.
Which cards are suitable for double-precision computation?
For FP64 workloads choose the A100, H100 or H200 — gaming cards have significantly lower double-precision performance.
Can I rent capacity for a university or a laboratory?
Yes, renting suits project work and requires no capital expenditure on equipment. Monthly rates are available for long-running projects.
Are MPI and job schedulers supported?
Yes, you get full root access over SSH and deploy the environment you need yourself, including MPI and schedulers.