---
title: Resources
description: Units for CPU, memory, GPU, storage and instances of applications, jobs and builds, and how requests and limits behave.
---

Kapten passes resources to Kubernetes as requests and limits, in Kubernetes units.

## CPU

CPU is set in **millicores**, written with an `m` suffix: `1000m` is one CPU core.

| Value | CPU |
| --- | --- |
| `250m` | A quarter of a core |
| `500m` | Half a core |
| `1000m` | One core |
| `2000m` | Two cores |

## Memory

Memory is set in **mebibytes**, written with an `Mi` suffix: `1024Mi` is 1 GiB.

| Value | Memory |
| --- | --- |
| `256Mi` | 256 MiB |
| `512Mi` | 512 MiB |
| `1024Mi` | 1 GiB |
| `2048Mi` | 2 GiB |

## Requests and limits

| Setting | Meaning |
| --- | --- |
| Request | What Kubernetes reserves for each instance. An instance is only scheduled on a node with this much free capacity. |
| Limit | The most an instance may use. Above its CPU limit an instance is throttled; above its memory limit it is stopped and restarted. |

Set the request to what the service uses under normal load and the limit to its peak. A limit far above the request lets one service starve others on the same node.

## GPU

Applications can request whole GPUs with a GPU count. The cluster must have GPU nodes and their device plugin installed.

## Storage

Persistent volumes are sized in whole **GiB**, and use a Kubernetes **storage class** of your cluster, such as `gp3` on EKS or `standard-rwo` on GKE. List the classes of your cluster with:

```bash
kubectl get storageclass
```

## Instances

Applications run between a **minimum** and a **maximum** number of instances, both at least `1`.

## Builds

Applications built from GitHub set the CPU and memory of their build separately from the application's own, in the same units, and a build timeout.
