Cuda docs suggest that more than one context on a card may slow down operations, so this might give us a slight win.