Author: Robert Hyatt
Date: 08:50:31 02/20/03
Go up one level in this thread
On February 20, 2003 at 09:18:35, Jeremiah Penery wrote: >On February 19, 2003 at 11:40:27, Robert Hyatt wrote: > >>On February 19, 2003 at 00:52:09, enrico carrisco wrote: >> >>>If done properly and tested for reliability -- what reasons do you speak of? >>>Most CPUs are purposely locked from higher than marked performance from the >>>manufacturer for marketing and other reasons -- both Intel and AMD. This, in no >>>way, means the CPU is incapable of such performance. >> >>The reason is _reliability_. As I explained in another post, circuits have a >>distinct settling >>time, which is what limits the clock frequency. Some circuits have varying >>settling times >>depending on the inputs. If you set the clock frequency too short, on occasion >>the circuit >>won't settle before the outputs are latched, and you get flakey results. Unless >>you do a >>complete and exhaustive test of all instructions, all inputs, you can't be sure >>you haven't >>stepped on this. I have debugged this problem in the past and don't _ever_ >>intend to debug >>it again. > >They don't make different circuits for different clock-speeds of the same chip. >As long as cooling is sufficient, it's pretty safe to overclock a lower-speed >processor to the same speed as the highest planned processor of the same core >design. The manufacturers generally don't push their processors to the limit, >so you can often go somewhat above that speed, even. We are just going to have to agree to disagree here. There are gate delays, and there are gate delays. This varies with voltage, resistance, inductance, capacitance, temperature, and probably other things I have omitted. Several things allow vendors to crank up the clock speed. Better transfer processes to make resistance more constant, which controls heat, which controls more resistance, which controls delays, etc. If you want to overclock, feel free. But you have to be aware that it is likely to bite you in very unexpected ways. If you build something from old TTL logic chips, you can look up their specs to get their gate delay. And when you add up all the gates in a single circuit, you _know_ that you had better keep the clock cycle time longer than the sum of the gate delays plus some constant depending on the wire path length. yes, it might run faster than that. Some of the time. But I'm not putting that in charge of a shuttle mission that has to be right _every_ time. :) > >It's not all that difficult to 'debug' any problems with overclocking. Run a >program like Prime95 for a while (catches errors MUCH faster than anything else >I've seen). If you get errors, clock down and see if they go away. Obviously, >errors are a sign of an unsafe overclock. It certainly doesn't mean that EVERY >overclock is unsafe. Otherwise, you'd better send back those 2.8GHz Xeons right >away, because they're nothing more than overclocked 2.5GHz ones (in a manner of >speaking). The problem is that the 'error signs' are not like red flags. It isn't easy to detect a single bit error in a floating point multiply, for example, if it only happens once every 10^6 multiplies with specific bit patterns. And that kind of failure has happened _many_ times...
This page took 0.01 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.