In article , keith wrote: [...]
They saved a lot of hardware in the denormalize logic at the cost of making the numbers less accurate. I think they also gained a bit of speed on the floating point add and subtract at the cost of making the sqrt() function take a little longer to do.
The denormalize could be made as just 1/16 muxes. When you want the initial guess for the sqrt() the wider range for the mantissa made your guesser more complex or less accurate.