Nvidia-smi fails on image: ubuntu-2004-cuda-11.4:202110-01

Following the documentation’s canonical example of how to use the GPU instances here: Using the GPU execution environment - CircleCI
I launched the following job:

version: 2.1

      image: ubuntu-2004-cuda-11.4:202110-01
      - run: nvidia-smi

but it failed with the error

NVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driver. Make sure that the latest NVIDIA driver is installed and running.
Exited with code exit status 9

and I don’t understand why? Doing couple of random ls -al, namely ls -al /usr/local/cuda-11.4 makes it seem like the drivers and the toolkit are installed, but the instance itself doesn’t have a GPU (meaning the error is real and not an artifact of a different one).

Since I heard back from the customer support on the issue I thought I’d update anyone else who runs into the question with the answer:

When specifying a GPU machine image, it is also necessary to include a special resource class in your config.yml file.

You can see a list of available resource classes on the following page:

Using the GPU execution environment - CircleCI

Using GPU-enabled resource classes also requires a paid Scale plan as mentioned on the following page:

CircleCI Resource Classes - CircleCI

An error message regarding the plan type should show in the UI if you try to add the gpu resource class to your job.

Thanks to the customer support for the quick response!