7 min read
How to train AI models without your own GPU
How to rent a GPU for training and fine-tuning models: how much memory you need, which card to pick, step by step, and how to pay per minute.
To train an AI model without your own graphics card, rent a GPU server by the minute: launch a server with the card you need, upload the data, train, download the weights and stop the server. On gpu.nz you do this from the browser or with the gpunz command; billing is per minute from a balance in rubles. The balance can be topped up with cryptocurrency through CryptoBot.
What training needs
Training needs video memory: the model weights, the gradients and the optimizer state have to fit on the card. For LoRA and QLoRA on 7–8B models, a 24 GB card is usually enough. Full training of large models needs 80 GB or more, or several cards at once.
Which graphics card to pick
- LoRA and QLoRA on 7–8B models: a 24 GB card, for example RTX 3090 or RTX 4090.
- Fine-tuning 13–34B models: a 48 GB card or more, or quantization to save memory.
- 70B models and training with long context: A100 or H100 with 80 GB, often several cards.
- Fine-tuning image models (SDXL, FLUX LoRA): 24 GB is usually enough.
The exact memory you need depends on the model size, the context length and the training method. The detailed estimate is in the guide «How much VRAM an LLM needs».
Step by step
- Check the training script on a small dataset locally, so you do not pay for bugs.
- Pick a card by memory in the catalog and launch a server with the PyTorch with Jupyter template (for images, Kohya SS).
- Add your SSH key before renting, then connect with the command from the server card.
- Check the card with nvidia-smi and run the training inside tmux or screen, so it survives a dropped connection.
- Save checkpoints to disk and download the weights before you stop the server.
- Stop the server: the GPU stops being billed. Delete the server if you no longer need the disk.
nvidia-smi
python train.py --epochs 3 --save-every 500What it costs
A minute costs the hourly rate divided by 60, and the price is fixed at launch. For example, two hours of training at 100 ₽ per hour cost 200 ₽. Interruptible servers are cheaper but can be paused, so save checkpoints on them. A stopped server is not billed for the GPU, but its disk is billed while it exists.
Payment
The balance is topped up with cryptocurrency through CryptoBot now: USDT, TON, BTC and other coins. Bank cards and SBP are coming soon.
Questions and answers
- Can I train a model on a rented GPU?
- Yes. Rent a server with enough video memory on gpu.nz, train the model and download the result. You pay only for the minutes the server runs.
- Which card is enough for LoRA?
- For LoRA on 7–8B models, 24 GB of memory is usually enough. For larger models, take a card with 48 GB or more.
- What happens to my data when I stop the server?
- The disk is kept and billed while it exists. Deleting the server deletes its disk too, so download the weights and results first.