Anything where you are doing massive and complex parallelism will require tonne of unsafe. Think co-operative thread groups on 120x4 = 480 core machines.