# Run GPU workloads

Run an application with GPU access on Northflank cloud or your own infrastructure. Services and jobs can use GPUs. Use a service for a continuous process and a job for work that ends.

> [!note] Requirements
>
> You will need the following to get started:
>
> - A container image that supports the intended GPU workload
> - The GPU type, memory, and resource requirements of the application
> - Access to a deployment location with the required GPU capacity

### 1. Choose where the workload runs

| Hosting choice | First step | Detailed guide |
| --- | --- | --- |
| Northflank cloud | Create a GPU-enabled project | [Deploy GPUs on Northflank cloud](https://northflank.com/docs/v1/application/gpu-workloads/deploy-gpus-on-northflank-cloud) |
| Your cloud | Prepare a cluster and GPU node pools | [Deploy GPUs in your own cloud](https://northflank.com/docs/v1/application/gpu-workloads/deploy-gpus-in-your-own-cloud) |

Review availability, requirements, and supported configuration in the selected guide before creating resources. The hosting paths have different GPU configuration constraints.

For your cloud, review the provider requirements and [node pool configuration](https://northflank.com/docs/v1/application/bring-your-own-cloud/deploy-and-scale-node-pools). Make sure that the cluster has GPU capacity before deploying the workload.

### 2. Configure the service or job

Use the selected hosting guide to choose the image and GPU resources. The shared configuration guidance is under [Compute → GPUs](https://northflank.com/docs/v1/application/gpu-workloads/configure-and-optimise-workloads-for-gpus).

Add the application's [runtime variables](https://northflank.com/docs/v1/application/run/inject-runtime-variables) and [secrets](https://northflank.com/docs/v1/application/secure/inject-secrets). If the image needs a different start command, [override its command or entrypoint](https://northflank.com/docs/v1/application/run/override-command-entrypoint).

For an HTTP service, [configure its port](https://northflank.com/docs/v1/application/network/configure-ports). A batch job does not need public HTTP access to perform its work.

### 3. Run a representative task

Examine the [container logs](https://northflank.com/docs/v1/application/observe/view-logs) and the application's own GPU diagnostics. Make sure that the application can use the selected device and complete its intended task.

Use [metrics](https://northflank.com/docs/v1/application/observe/view-metrics) to review resource use. For your cloud, also review node pool capacity and cluster operation in the provider guide.

### Next steps

For model serving, continue with [Host a model on GPUs](host-a-model-on-gpus). For tasks that end or use a schedule, read [Run background tasks](run-background-tasks).