Start
GPUs & pricing
Slices, whole cards, spot and what you pay.
Every GPU on Flint runs in a container on a real NVIDIA card. You choose how much of a card you need.
Slices and whole cards
| Option | What you get |
|---|---|
| Slice (e.g. 2 GB or 4 GB) | A share of one card: a hard GPU-memory limit, a guaranteed share of the card's compute, and your own CPU, RAM and disk. Other customers may use the rest of the card. |
| Whole card | The entire GPU, all its memory and compute. Nobody else runs on it. |
Memory limits are hard. If your program asks for more GPU memory than the slice has, it gets a CUDA out-of-memory error, exactly as on a smaller card. nvidia-smi inside the GPU shows your slice's limit.
Compute share is the fraction of the card's processing power you're guaranteed when the card is busy. A 2 GB slice of an RTX 4060 is sold as about 25% of the card, a 4 GB slice as about 50%. When neighbours are idle you may get more.
Customers are isolated from each other: separate containers, separate networks, and GPU memory is wiped by the driver before it's handed to the next customer.
On-demand and spot
- On-demand slices get their compute share whenever they need it.
- Spot slices cost about half. When an on-demand customer on the same card is busy, a spot slice's GPU work is held back until the card is free again, then it continues at full speed. Your process keeps running and your memory is kept; it just pauses.
Spot suits batch jobs, experiments and fine-tunes that can wait. Use on-demand for anything interactive or time-sensitive.
What you pay
Prices are per hour, shown in the launcher and on each GPU. Current examples on an RTX 4060:
| GPU | Price |
|---|---|
| 2 GB slice | US$0.09/h |
| 4 GB slice | US$0.15/h |
| Whole card (8 GB) | US$0.29/h |
| 2 GB slice · spot | US$0.05/h |
| 4 GB slice · spot | US$0.08/h |
- Running: billed by the minute. Launching sets aside 30 minutes of credit; minutes you don't use stay in your balance.
- Paused: a reduced rate (shown when you pause) keeps the GPU reserved for you.
- Terminated: no GPU charges.
- Shared volumes: US$0.10 per GB-month, whether or not they're attached. See Storage.
The launcher shows your total per hour and how long your balance lasts at that rate before you launch.
Every GPU includes
- Ubuntu with CUDA, PyTorch and JupyterLab
- Root access over SSH, a browser terminal and JupyterLab
- A
/workspacedisk that survives pause and resume - Optional exposed ports and shared volumes
Flint documentation