#TIL -- that in linux, if you use an Nvidia GPU for compute (i.e., not graphics) the driver could be offloaded on idle times. There is a persistence mode which is supplemented by a persistenced service to prevent this.
*I guess it depends on which linux flavor and which nvidia driver, but seems common
If the driver is being offloaded on idle times, there is a cold start every time you need the compute. Good to know for headless GPU compute servers

