Well, I mean it is measuring how reliably you can get a GPU instance. But it certainly isn't the overall reliability. And depending on your workflow, it might

Well, I mean it is measuring how reliably you can get a GPU instance. But it certainly isn't the overall reliability. And depending on your workflow, it might not even be a very interesting measure. I would be more interested in seeing a comparison of how long regular non-GPU instances can run without having to be rebooted, and maybe how long it takes to allocate a regular VM.