Allow me to attempt explaining it: A program is a collection of one or more "threads" that you can think of as semi-independent little programs sharing memory. Most often, one is assigned per GPU you have, to handle giving that specific GPU work, and getting results from it. When you ask a GPU to do something - you are actually "calling" (or running code from, and waiting for it to tell you it finished and what went down) code from the driver. Keeping it simple - AMD develops the driver, which is buried deep in the OS and controls the hardware for you - so you can say "Run this shit and tell me what nonces are good" without having to care too much about the details. If a GPU has hung... it has hung inside the code that is deep in the shit that the miner dev usually doesn't even think about - let alone control. That thread is now "stuck", and will remain so until the entire program terminates. You may notice how when a GPU goes "DEAD" in SGMiner, the others continue to mine... this is why.

So, everything continues despite that thread's fucked status until somebody or something tells the whole miner program to quit... and therefore, all its threads must quit... oh shit. The OS can't actually terminate it properly - it's stuck in low-level code that happens to often be a core, low-level part of the OS itself... in short - we be fucked. Hard reboot to bring the GPU back up, in short.