What Server Do You Need for Machine Learning and AI

Find out where a CPU is enough and where a GPU is required — then check the build on your own data during the 5-hour test period

Match a configuration
Choosing a build for ML and AI

Choosing a server for machine learning starts with one question: do you actually need a graphics card. A CPU is enough for classical machine learning, data preparation and inference of small models: a 7-billion-parameter model takes 4–5 GB in 4-bit quantisation and 7–8 GB in 8-bit, and with context plus the operating system you plan for 16 GB of RAM or more. A GPU is required for training neural networks and running generative models — that configuration is not part of the plans and is assembled for the task. Below are builds by scenario; once you have chosen, order a VPS for AI/ML.

What such a server costs

Plans without a graphics card — for data, classical ML and inference. Test period of 5 hours before payment.

5.49 mo

Standard 1

4 vCPU · AMD EPYC
4 GB RAM
50 GB NVMe
Unlimited traffic
IPv4 · KVM
24/7 support
6.99 mo

Standard 2

4 vCPU · AMD EPYC
6 GB RAM
60 GB NVMe
Unlimited traffic
IPv4 · KVM
24/7 support
8.99 mo

Standard 3

6 vCPU · AMD EPYC
8 GB RAM
80 GB NVMe
Unlimited traffic
IPv4 · KVM
24/7 support

A build for a typical task

Memory and core benchmarks for workloads that run on the processor.

Task Configuration Plan Price, €/mo
Study projects, scikit-learn on small data 6 vCPU · 8 GB · 80 GB Standard 3 €8.99 Order now
Jupyter and dataset preparation 8 vCPU · 8 GB · 100 GB Standard 4 €10.99 Order now
Inference of a quantised 7B model 10 vCPU · 16 GB · 160 GB Power 1 €17.99 Order now
Gradient boosting, ETL on large tables 16 vCPU · 24 GB · 240 GB Power 3 €27.99 Order now
Datasets of hundreds of GB, heavy preprocessing 40 vCPU · 64 GB · 700 GB Max 3 €84.99 Order now

Benchmarks for computation on the processor. Training neural networks requires a build with a graphics card: it is not part of these plans and is matched individually to the model and the data volume. Prices are for monthly billing from our price list — 10 % lower for six months, 18 % lower for a year.

1

Memory for the model and context

Add the model's weight in your chosen quantisation to the context and the system: 7B in 4 bits is 4–5 GB, and 16 GB of RAM is the working minimum.

2

Cores for the data

Preprocessing, ETL and gradient boosting parallelise well — here 16–40 vCPU help, not a graphics card.

3

A GPU is for training

Training neural networks and generative models needs a graphics card. We assemble that build per task — see the GPU server comparison.

How to match a server to your task

Three steps — from the model's requirements to a measurement on your own data.

1

Identify the type of workload

Training a neural network means a GPU. Inference, classical ML and data processing run on the processor.

2

Calculate the memory

Add the model's weight in the quantisation you need, the batch or context size and 2 GB for the system — that is your RAM minimum.

3

Check it on your own data

The VPS is provisioned in test mode for 5 hours: set up the environment, run an epoch or a batch of inference and decide before paying.

The build is chosen — time to order

The build is chosen — time to order

Plans and the ordering procedure are on the VPS for AI/ML page. If you need a graphics card, start with the GPU server comparison: it explains what to check besides the card model.

If you already know the tool, start from its own page: PyTorch, TensorFlow, Jupyter or Stable Diffusion.

Go to the VPS order page

Frequently asked questions

Didn't find your answer?
Write to us via the contact page form.

A CPU is enough for classical ML, data preparation and inference of small models. A GPU is needed for training neural networks and generative models — we assemble that build per task.

About 4–5 GB in 4-bit quantisation and 7–8 GB in 8-bit. With context and the system, plan for 16 GB of RAM or more — that is Power 1 at €17.99/mo.

On the server's own NVMe: 160 GB in Power 1, 240 GB in Power 3, up to 700 GB in Max 3. If you have more data, write to support and we will attach an additional disk.

Any of them: PyTorch, TensorFlow, scikit-learn, JAX, conda environments and whichever driver versions you need. Root access is full and we do not interfere with your software stack.

ETL, preprocessing and gradient boosting parallelise well: 16 vCPU (Power 3, €27.99/mo) are noticeably faster than 8 on the same tables.

Yes, Jupyter installs on the server and opens in a browser over an SSH tunnel or a reverse proxy. Your notebooks and data stay on your own server.

The price depends on the card model and the amount of video memory, so it is quoted per task. Describe the model and the data volume and we will assemble a build and price it.

On the VPS for AI/ML page: plans, OS images and checkout. The server is provisioned in test mode for 5 hours — you pay after checking it.