Just attach to a 2nd GPU project with a resource share of zero. Back-fills only when needed. No fiddling or monitoring needed. Works great!