[comp.arch] error magnitude in i860 divide - reread the question.

jeff@oakhill.sps.mot.com (Jeff Enderwick) (06/30/91)

Maybe I wasn't clear in my original posting about this subject. I understand
how iterative reciprocal approximation works. I realize that the error 
analysis from a NA standpoint should be simple. My question is this:

	I used the simple f(x,y) = fm(x,y) * ( 1 + e ) analysis method
	and I found 3 bits of worst case error. The i860 manual that I 
	have says 2 bits. I'm looking for someone who can tell me 
	why/how I screwed up in my analysis. Maybe there is something
	inherent in the algorithm that makes the standard method overkill.

	An Intel i860 apps number would also be appreciated. 

To reduce wasted bandwith:
	1) I'm not calling Intel a liar. I hope that 2 bits *is* right.
	2) forget #2. It'll just start a flamewar.


	Jeff Enderwick
	Motorola DSP Compilers.

	jeff@otis.sps.mot.com