[perfmode]

deep learning uses low precision floats

sometimes as few as 8 bits are needed

[jadedhacker]

I think gen 1 or gen 2 of the TPU explicitly supported short ints.