NONMEM Users Network Archive

Hosted by Cognigen

Extremely geeky question

From: Mark Sale - Next Level Solutions <mark>
Date: Thu, 13 Nov 2008 07:55:32 -0700
  Most modern computers use a system cal= led ECC (Error checking/correction) wherein each memory location has an ext= ra bit (the parity bit) that is used for checking whether the data location= is "correct".  Interestingly, data bits in memory can actually change= randomly, due to electro magnetic radiation - even cosmic rays.  Some= computers do not have ECC, most recently, the Pentium series did not. = ; (note that this is not the division problem that some early Pentium CPUs = had, this is a random event).  The Core 2 duo has ECC.
  If th= ere were a an random data change, NONMEM could give different results from = the same run.  It might very well recover, since it reads the data set= from the disc source (which would not change, only stuff loaded into DRAM = could change) for each iteration, and presumably, the correct values for da= ta, and all the variables internal to NONMEM could be fixed next iteration = (the Hessian might take a while to recover, and there are certainly other v= ariables that changes could not recovered from at all).  
&nb= sp; This is relevant because the next generation of Intel CPU (Core i7, to = be released next week) do not have ECC.  Actually it is the mother boa= rds for the Core i7 that do not support ECC.  A leader in the field (D= aniel Bernstein) contents that in scientific computing you don't need ECC u= nless you actually care that the answer is correct (or something like that)= .
  Here is the question.  Has anyone ever seen apparently ran= dom - non reproducible results from NONMEM.  That is, run the same dat= a set, same control file, same hardware, same compiler and options and got = a different result. If so, any idea what hardware it was on?
<= br> 

<= /div>

(image/png attachment: left.letterhead)

Received on Thu Nov 13 2008 - 09:55:32 EST

The NONMEM Users Network is maintained by ICON plc. Requests to subscribe to the network should be sent to:

Once subscribed, you may contribute to the discussion by emailing: