The problem is that the driver might be in a corrupt state, and the miner can't shut down cleanly. On my own rigs, I need to reboot anyway in 99% of these cases.

That said, my experience is that the win driver(s) recover somewhat better for Polaris cards. If you know you'll get a clean shutdown and always can just restart the miner, I would define a watchdog.bat file in your miner dir that (1) sleeps for 20 secs or so, (2) runs your normal startup procedure (resets clocks for all gpus, start the miner, etc). Last, you also need to add the command line arg --watchdog_script when starting the miner.

The procedure when a dead GPU is detected is that the miner will both execute your watchdog script and then shut down. The 20 sec sleep in your watchdog should catch any unexpected wait times during the miner exit.