AI Score: how we rate GPUs
Oct 29, 2026 · @ates
What AI Score is
AI Score is a single number that shows how good a GPU is for AI workloads: running language models, generating images, training neural networks. The higher the number, the faster the card handles this kind of work.
AI Score is shown on every machine's card in the marketplace. You can filter and sort offers by it to quickly find a suitable machine.
What it measures
The test checks three things that determine how fast real AI workloads run:
What is tested | Why it matters |
|---|---|
Compute power | How fast models train and long prompts are processed |
Video memory speed | How fast a language model writes its answer |
Training a small neural network | How the card performs on a complete task where both compute and memory matter |
The final number weighs all three parts equally, so one strength won't hide a weakness. Video memory size is not part of AI Score: it is listed separately in the machine's specs.
How to read the number
Reference point: a single GeForce RTX 4090 scores about 60,000. The scale is linear: a card twice as fast scores twice as much, a card twice as slow scores half as much.
If a machine has several GPUs, their AI Scores add up. A machine with four RTX 4090s scores about 240,000. Keep in mind that not every workload can use several cards at once: for those, the AI Score of a single card matters more.
Small differences between identical cards are normal: the result depends on cooling, the specific manufacturer's design and power settings.
When the test runs
The test runs automatically once, after the machine passes verification, together with the other tests (disk, network, Blender). The machine owner can rerun the test in their dashboard, for example after replacing a GPU or updating the driver.
On a modern card the test takes about 15 seconds, longer on older ones. Cards are tested one at a time. On the first run the machine downloads the test first, which depends on its internet speed.
While the test runs, the machine is busy and unavailable for rent, and background jobs are paused. It returns to work right after the test. The test never interrupts a rental: a busy machine is not tested.
Supported GPUs
AI Score is measured for NVIDIA GPUs from the GTX 900 series up to the newest ones, including RTX 50 and professional cards (RTX A, RTX Ada, RTX PRO, A100, H100).
- NVIDIA driver 525 or newer is required.
- RTX 50 cards and other Blackwell-architecture cards require driver 570 or newer.
- AMD and Intel GPUs are not supported yet, and their AI Score is 0. This doesn't mean the card doesn't work: the test simply isn't run for them.
Why AI Score can be 0
Zero means the test could not be run. The reason is shown in the machine's test results in the dashboard.
Reason | What to do |
|---|---|
The GPU is not NVIDIA, or there is none | Nothing: the test isn't run for such machines |
The GPU is too old | Cards older than the GTX 900 series are not supported |
Outdated driver | Update the NVIDIA driver and rerun the test |
The GPU is not visible to the system | Check that the driver is installed and the card is detected, then rerun the test |
Not enough video memory | Close programs that are using the GPU and rerun the test |
The machine went offline or was busy during the test | Rerun the test when the machine is free and online |
If a machine has several cards and at least one fails the test, the AI Score of the whole machine is 0. The results show which card failed.