Generally speaking GPUs are already very good at running with float32s, usually much better than they are at using float64s in fact. The big advantages of usin

Generally speaking GPUs are already very good at running with float32s, usually much better than they are at using float64s in fact. The big advantages of using an ASIC are mostly on the storage side but they also allow you to get away with non IEEE floating point numbers that don't necessarily implement subnormals, NaN, etc.