Lionel Lacassagne
Laboratoire d'Informatique de Paris 6 (LIP6) - Universite Pierre et Marie Curie (UPMC)
System on Chip Department (SoC)
Home
origine
format
algorithm
RISC processor
embedded processor
FPGA & SoC
F16 Format
sometimes size matters...
Advantages
F16 format has several advantages:
automatic vectorisation like 32-bit FP
smaller cache footprints than 32-bit FP
better precision than 16-bit integer
two kinds of format are possible:
IEEE 754 Format
for 16-bit wide, IEEE format is
1-bit sign
5-bit exponent with a bias of 15
10-bit mantissa
Dynamic range is 65000 to 2^-8
custom Format
parameters can be tuned:
still a 1-bit sign
mantissa could be smaller than 10 bits
exponent bias can be smaller to take into account "big numbers" (close to 2^32)
Dynamic range now depends on exponent, and precision depends on mantissa size
Big Brother is watching you
just smile ;-)