5 ms·
You already have that; it's more formally called IEEE 754 single precision. Requires double the memory bandwidth (bad) for extra precision that back propagatio
by ndesaulniers 7y ago
You already have that; it's more formally called IEEE 754 single precision. Requires double the memory bandwidth (bad) for extra precision that back propagation would have corrected for anyways.
- ScottBurson 7y agoSeems like BFloat32 would have the same exponent range as Float64 (double precision), and a correspondingly shorter significand.