Conventional numbers are easy to add/subtract and logs are easy to multiply/divide — I wonder if there is a numeric representation that could support more efficient hardware fused multiply-adds for matrix multiplies.
Numeric representation for efficient fused multiply-add hardware
By
–