What is on-demand
GPU compute?

What a GPU does that a CPU does not

A GPU runs many similar calculations in parallel. That suits the matrix operations in machine learning, image rendering and some simulations. A CPU handles a broader mix of tasks and is often better suited to sequential work, application logic and coordinating the job.

A GPU is useful when your software can use it. Adding one does not automatically make every program faster. Check your application’s GPU support and required drivers before renting.

On-demand, reserved, spot or buying?

On-demand

Rent when you need it and release it when you finish. Useful for changing workloads or a project with an uncertain duration. Check billing increments and availability.

Reserved

Commit to capacity or usage for an agreed period. This may suit predictable work, but the payment and cancellation terms matter if your plans change.

Spot or interruptible

Use capacity the provider can reclaim. Suitable for work that can save checkpoints and restart. An interruption can delay results or require repeated work.

Buying hardware

Own and manage the machine. Consider ongoing usage alongside power, cooling, maintenance and the time needed to run it.

Size the job: memory, throughput, time

Memory first

The model, scene or working data must fit in memory, along with intermediate results and software overhead. GPU memory and system RAM are separate constraints. Several GPUs do not automatically behave like one larger pool of memory; your software must support splitting the work.

Then throughput

Once it fits, consider how much work must finish per unit of time. Software support, numeric precision, CPU resources, storage speed and data transfer can all affect throughput. Describe the actual job rather than choosing from a hardware name alone.

Then time

Share your deadline and expected runtime. Include time to upload data, install dependencies, run a small check and retrieve results. If you do not know the runtime, ask whether a short trial can help size the full job.

Virtual workstations

A virtual workstation is a remote computer you use interactively, often through a remote desktop. It can suit editing, 3D work, development or other tasks that need more resources than your own device.

Tell the provider your operating system, applications, licence requirements and where you will connect from. Network latency and display quality matter for interactive work. Ask how files persist, how access is secured and how to stop the machine when you finish.

The costs beyond the machine

• Storage: disks, snapshots and saved data may remain billable after a machine stops.
• Egress: moving results out of a provider can incur data transfer charges.
• Idle time: a running machine can keep accruing charges while you are away or waiting for an upload.
• Setup and software: licences, support or configuration work may be separate from compute.

Ask what is included, what stays active after stopping and how to remove resources when the project ends. For Compute On Call, pricing is quoted in your reply.

A checklist to send a provider

• The task, software and versions you need to run.
• The model, scene or dataset size and known memory requirements.
• GPU needs, CPU cores and system RAM, if known.
• Storage size, data location and upload or download needs.
• Your start date, deadline and expected duration.
• SSH, Jupyter notebook or remote desktop access.
• Privacy, hosting region or compliance requirements.
• Whether the job can tolerate interruptions and resume from checkpoints.
• Budget constraints and who should receive the quote.

Leave unknown specs open. A clear description of the work is a useful starting point.

Common questions