Ubuntu + CUDA is a GPU server with the NVIDIA driver and CUDA already installed, and everything else is up to you. You log in over SSH as root and work as on any Linux server. It suits your own code, builds and services that need a GPU but not a ready-made interface.
In short
Ubuntu + CUDA runs on cards with 8 GB or more. The most affordable suitable card right now is the RTX 3060 12 GB from 6.95 ₽ per hour, billed by the minute.
Updated 13 October 2026
Ubuntu + CUDA does not limit the choice of card, so take the one your task needs. For inference and experiments with 7–8B models 12–16 GB is enough, for example an RTX 3060 12 GB. LoRA training of such models usually needs 24 GB, as on an RTX 4090 24 GB. Large models and distributed jobs need 80 GB or more, such as an A100 or H100.
Prices and availability come from the live catalog and update automatically.
The NVIDIA driver and CUDA come with the image. Check them with one command: the header of the output shows the card and the CUDA version the driver supports.
nvidia-smiPick PyTorch and other framework builds for that CUDA version or an earlier one. If your job can tolerate interruptions, interruptible servers cost less, but the host can take such a server back, so save your results often.
Install system packages with apt, and Python environments with venv or conda. For PyTorch take the command for your CUDA version: for example, the cu124 index matches CUDA 12.4.
apt update && apt install -y git tmux
python3 -m venv ~/venv && source ~/venv/bin/activate
pip install torch --index-url https://download.pytorch.org/whl/cu124Run long jobs in tmux so they survive a dropped connection, and copy files from your computer with scp. The guide SSH on a GPU server explains the connection in more detail.
ssh-keygen -t ed25519 # on your computer
ssh -p <port> root@<address> # the command from the server page
scp -P <port> ./dataset.zip root@<address>:/root/
tmux new -s train # a session for a long jobThe GPU is billed while the server runs, and the disk is billed for as long as the server exists. Stopping the server between sessions and deleting servers you no longer need noticeably cut costs: see the guide how to save on GPU rental.