Does anyone know how they can detect hardware defects like this? This sounds like an incredibly hard problem. And I don’t see how they can do this without impacting performance significantly.

If the crash is isolated (no other reports) and flipping one bit in the crashing pointer value would make the pointer valid, it's assumed to be a bitflip. This obviously will only catch a minor portion of bitflips, i.e. any image or video data with bitflips wouldn't crash.

From what he's saying they run an actual memory test after a crash, too.