# RTX 3090 cloud Europe: budget 24 GB GPUs for EU experiments URL: https://cloudj.net/blog/rtx-3090-europe Published: 2026-11-12 ยท Updated: 2026-11-12 > Use RTX 3090 cloud Europe hosts for hobby and small-team ML: EU codes including Eastern Europe, and a per-minute routine for short, cheap experiments. An RTX 3090 cloud Europe setup is one of the most practical ways for a hobbyist or a small team to get 24 GB of VRAM without buying a card. The 3090 is an older consumer GPU, but its memory is the same size as a 4090's, and for many experiments memory matters more than speed. On Cloudj you search for it by typing EU country codes, including Eastern European ones, into the Country field of the GPU deploy page. Because billing is per minute, the real skill is running short sessions that start fast, do the work and end cleanly. This article covers both. For all European codes and how they compare, see the [GPU cloud in Europe guide](https://cloudj.net/blog/gpu-cloud-europe-guide). ## Why the 3090 still makes sense The [RTX 3090](https://cloudj.net/gpu-cloud/rtx-3090) has 24 GB of VRAM, and so does the 3090 Ti. That is enough for: - running a quantized mid-size open model to test prompts, RAG pipelines or agents; - LoRA or QLoRA fine-tuning of small and mid-size models with modest batch sizes; - diffusion image generation and training small add-ons; - classic computer vision training on your own datasets. A newer card such as the [RTX 4090](https://cloudj.net/gpu-cloud/rtx-4090) has the same memory and is faster, so jobs finish sooner. Whether that saves money depends on the live prices at the time and on how long your job really runs. Compare both on [GPU cloud](https://cloudj.net/gpu-cloud) before you choose. Both are consumer cards, open to new accounts. If you lean towards the newer card, the [RTX 4090 in Europe guide](https://cloudj.net/blog/rtx-4090-europe) covers its search. ## Search with EU codes, east included Open the [deploy page with the RTX 3090](https://app.cloudj.net/deploy/gpu?model=rtx-3090) and enter a broad set of codes: ```text DE, PL, RO, NL, CZ ``` This mixes central and western Europe with Eastern European countries. Each row still shows its country, so you can prefer one if it matters. Eastern Europe is worth including for a simple reason: a wider search gives the price sort more machines to choose from. It is not a promise that any country is cheaper; prices and stock differ by country and change during the day. The [Eastern Europe GPU guide](https://cloudj.net/blog/gpu-cloud-eastern-europe) covers those codes in more depth. If you are new to the fields, the [country filter walkthrough](https://cloudj.net/blog/filter-gpu-by-country-console) explains each one. If nothing is listed, add `FR`, `SE` or `GB`, or switch the model to the 3090 Ti or the 4090. ## Which columns matter for experiments For short experiments, the ranking is different from a production service: | Column | Weight for experiments | Reason | |---|---|---| | Total price | High | You pay per minute; lower is better for short runs | | Download speed | High | Pulling weights and datasets is often the slowest step | | Bandwidth price per TB down | Medium | Large downloads add up, uploads are usually small | | Reliability | Medium | A crash costs you a rerun, not customers | | Max duration | Low | Experiments last hours, not weeks | A host with fast downloads can save more time than a faster GPU, because each new session starts by pulling something. ## A per-minute routine for short sessions The cost of an experiment is mostly the time the instance runs. The routine below keeps that time close to the time you actually use the GPU. 1. **Prepare before you deploy.** Have your script, requirements and dataset location ready on your laptop or in a repository. 2. **Deploy with the right disk.** Disk size cannot be changed later. Leave room for weights, the environment and outputs. 3. **Connect and check the GPU.** Use the SSH command from the Connection tab, then run `nvidia-smi`. A newly added SSH key can take up to three minutes to be accepted. 4. **Pull and run in one go.** Start the job inside `tmux` so it survives a dropped connection. 5. **Copy results out as soon as they exist.** Push outputs to your own storage with `rsync`. 6. **End the session deliberately.** Destroy if you are done, stop only if you will come back soon. ```bash ssh -p root@ tmux new -s exp git clone && cd && pip install -r requirements.txt python train.py --config configs/lora-small.yaml rsync -avP -e "ssh -p " root@:/workspace/outputs/ ./outputs/ ``` The last command runs from your own machine. ## Stop or destroy at the end? Stopping keeps the disk and charges storage only, but it releases the GPU. Someone else may rent it, and starting again may show "Waiting for GPU" until it is free. For experiments, that often means destroying is the better habit: save outputs, destroy, and deploy fresh next time. The [stop vs destroy guide](https://cloudj.net/blog/stop-vs-destroy-gpu-instance) has a checklist for each. When you destroy, the unused part of your deposit returns to your balance, possibly after up to 72 hours while bandwidth charges settle. ## Limits worth knowing - The 3090 is a consumer card in an independent host with no hardware SLA. Keep copies of anything you care about. - There is no graphical desktop; everything runs headless over SSH. - Respect the licence of each model you download, especially for anything you plan to publish. ## FAQ ### Is a 3090 good enough for LoRA fine-tuning? For small and mid-size models with LoRA or QLoRA, usually yes. Check memory with `nvidia-smi` during a short trial run. ### Are Eastern European hosts cheaper? Not necessarily. Prices differ by machine and change often. Include those codes to widen the list, then let the price sort decide. ### Can I pause an experiment overnight? You can stop the instance, but the GPU is released and restarting may wait. If you need the GPU in the morning, save outputs, destroy, and deploy again. ### How am I billed? Per minute from your prepaid balance, with the price locked at deploy. ## Try a session Open the [deploy page with the RTX 3090](https://app.cloudj.net/deploy/gpu?model=rtx-3090), enter `DE, PL, RO, NL, CZ` and pick a row with good download speed. Run your experiment in one focused session, copy results out, and destroy the instance when you are done.