It seems like you simply restated your question without taking into account anything I said.
I have no idea what that is. A google search turns up nothing. It’s not part of CUDA C++ from what I can tell.
The largest range you can have is -2048 to +2048. Maybe with some fiddling you could remap that to 0 to 4096. 4096 is 2^12. So no, you cannot have a range of 2^18, without gaps. You don’t get to use all 19 bits (or 16 bits in the case of FP16) when storing an integer in a floating-point (TF32 or FP16) quantity. Even if you think you could with some grand bit-mapping scheme, you won’t be able to pump any such scheme through the TC unit with sensible results. The range limit for direct integer storage is indicated by the number of mantissa bits, not anything else.