Author: Robert Hyatt
Date: 16:08:18 07/03/03
Go up one level in this thread
On July 03, 2003 at 13:25:31, Keith Evans wrote: >On July 02, 2003 at 17:51:47, Aaron Gordon wrote: > >>When trying to find calculations on a cpu you never, ever calculate by the chips >>"rated" speed, the core is almost always much better. Take the latest >>hand-picked 1700+ (1.46GHz) chips. Lets say after 24-48 hours of testing we find >>the core can do 2440MHz at 1.75v, completely stable via Prime95/BurnK7 for hours >>and hours at 60C (via Standard heatsink and fan running low rpms, to raise cpu >>temp to 60C intentionally). > >I think that there's one fundamental place where we disagree and it occurs early >in the process. You make the assumption that running Prime95/BurnK7 is >sufficient to verify that a chip is operating correctly when run out of >specification. I question this. The documentation for BurnK7 has a disclaimer >stating that it alone is not sufficient. How can we verify your fundamental >assumption? You can test every instruction with every possible set of operands, over every possible set of temperature/voltage variations. :) There is _no_ other way to guarantee that a processor works correctly, whether it is overclocked or underclocked. It is similar to testing a disk drive with a surface test. The more comprehensive the test, the more time required. And the more patterns get read/written. If you go back in time to somewhere around 1990, you will find stories about my running on a Cray C90 where _one_ processor dropped one bit in one strange case. And it caused some very interesting problems. Over the course of the game it went from one bit to 2 bits and kept getting worse, which would cause the program and/or O/S to crash. We finally had to drop back to one CPU, and we could run for a good while until we would randomly hit "the cpu" and go south again. That machine had run for 30 days with _zero_ errors and was ready to ship. Testing is _tough_.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.