RE: Defective PENTIUM FPU's

[THPierce (- at -) rohmhaas.com (Tom Pierce)]

 Previously David J. Heisterberg (djh (- at -) ccl.net)  wrote (in response to
 John
 McKelvey):
 >>NOW WHAT? COMPUTE STATISTICAL PROBABILITIES OF ERRORS IN COMPUTED
 RESULTS?
 >Tim Coe at Vitesse Semiconductor has done just that.  He's posted a
 >very detailed description of the problem in the comp.arch.arithmetic
 >news group.  For the likelyhood of running into this bug he says
 >  that needs clearing is that this is an extended precision problem.  This
 >  bug hits between 50 and 2000 single precision dividend divisor pairs (out
 >  of a total of 64 trillion.)
 >His analysis shows that the probability of getting a relative error
 >greater than 1e-5 is 3e-12, of less than 1e-5, 6e-12.  As to when this
 >error occurs
 >  Examination of the above divide failures reveals that both the dividend
 >  and divisor are integers minus small deltas.  Also notable is the induced
 >  error is roughly delta^(2/3).  The integers in the divisors are actually
 Hmmm - integer minus small delta's ala x-y=4195835-3145727=1050108,
 a large delta, and the error is 256 -> delta^2/3 =4096 not 1,050,108.
 I am unhappy with the processor, and now will not use it for calculations...
 Maybe I should send it back... Where is that warranty card...
 --
 Sincerely,
 Tom Pierce Ph.D.