# GPU capacity planner

Use a measured workload in GPU-hours to estimate the allocation needed before your deadline. This assumes the workload scales across GPUs.

Use the tool on the [web page](https://gpuclouds.fyi/tools/gpu-capacity-planner). It works without JavaScript.

## How to use it

1. Enter measured workload in gpu-hours.
2. Enter available gpus.
3. Enter parallel efficiency as a percent.
4. Enter hours until deadline.
5. Read the result.

## How it works

Elapsed hours = measured GPU-hours / (GPU count × parallel efficiency). This is a planning model, not a vendor benchmark. A zero result with missing inputs is not a valid runtime.

Measure efficiency with the same model, batch, precision and interconnect. Do not assume linear scaling. Include data loading and checkpoint overhead in your measured work.

Confirm quota and device allocation before scheduling. Kubernetes GPU scheduling documentation describes device plugins and GPU resource requests: https://kubernetes.io/docs/tasks/manage-gpus/scheduling-gpus/ , fetched 2026-10-29.

> Planning only. Use a measured runtime and a current quote. Taxes and unentered fees are excluded.

## Questions

### Is this a capacity guarantee?

No. Confirm availability and the full invoice terms with the provider.