> ## Documentation Index
> Fetch the complete documentation index at: https://docs.hivenet.com/llms.txt
> Use this file to discover all available pages before exploring further.

# GPU types and sizes

> Pick a GPU instance on Compute with fixed vCPU, RAM, disk, and bandwidth. Choose RTX 4090 or RTX 5090 based on your model needs.

Hivenet GPU instances pair NVIDIA cards with a fixed amount of vCPU, ECC RAM, and disk so you can launch fast without tuning every knob.

### Available GPU families

* **NVIDIA RTX 4090** — 24 GB of GDDR6X memory per GPU
* **NVIDIA RTX 5090** — 32 GB of GDDR7 per GPU

<Note>
  “RAM” in the tables refers to system memory (ECC).
</Note>

### Sizes and specs

| Size    | GPUs                | vCPU | RAM (ECC) | Disk space | Bandwidth |
| ------- | ------------------- | ---- | --------- | ---------- | --------- |
| X-Small | 1 × RTX 4090 / 5090 | 8    | 48 GB     | 250 GB     | 1 Gb/s    |
| Small   | 2 × RTX 4090 / 5090 | 16   | 96 GB     | 500 GB     | 1 Gb/s    |
| Medium  | 4 × RTX 4090 / 5090 | 32   | 192 GB    | 1 TB       | 1 Gb/s    |
| Large   | 8 × RTX 4090 / 5090 | 64   | 384 GB    | 2 TB       | 1 Gb/s    |

### Quick tips

* **VRAM vs RAM:** VRAM affects maximum model size per GPU; system RAM supports your runtime and dataset loaders.
* **Throughput:** Scale from 1 → 8 GPUs for higher parallelism.
* **CPU host:** Nodes use **AMD EPYC 7713** processors.