In development

How to choose a GPU server

A GPU server is worth it when the work is large, parallel math, such as running or training a model. Before choosing one, check whether the model fits, whether the software matches, and whether a CPU would do.

In development. Nothing on this page is on sale yet, and there is no account to create. The tools linked here are free and work today.

What we are building

ANAROQ does not run GPU capacity, and a GPU product is not part of the current plan. This page is a guide, not an announcement.

If GPU compute is added later, it will sit under the same layer as everything else: limits, cost before action and an audit trail.

How to choose

What to check before you commit, and which tool on this site answers each question.

  1. Ask whether you need a GPU

    Small models can answer on a CPU, more slowly. Try that first: the Stack Sizer estimates Open WebUI with Ollama on CPU only, so you can see what a plain server gives you.

    Tool: Stack Sizer

  2. Fit the model in GPU memory

    The model's weights and the context it keeps while answering have to fit in GPU memory. Check the model card; a quantized version needs less memory at some loss of quality.

    No tool for this step: ask the provider and read their terms.

  3. Match the driver and runtime

    Check the GPU driver, the CUDA or ROCm version and the container image you plan to use against each other before you rent anything.

    No tool for this step: ask the provider and read their terms.

  4. Prepare it like any server

    A GPU server still needs a user, SSH keys, a firewall and updates, and the apps still need files you can rebuild from.

    Tool: cloud-init Generator for Ubuntu 24.04

  5. Run the stack from files

    Write the model server and its interface as containers with pinned images, so a new machine runs exactly the same thing.

    Tool: Docker Compose Generator

  6. Measure with your own prompts

    Throughput depends on the model, the prompt length and how many requests run at once. Measure with your real workload, and stop the machine when it is idle.

    No tool for this step: ask the provider and read their terms.

What this page does not cover

This page states no GPU figures. Memory sizes, speeds and prices change often, and the site publishes such numbers only with a source and a date.