There are plenty of issues with current methods of comparing temperature data between processors, and we have been hard at work trying to come up with something that gives us better and more comparable results than the temperature data reported by different processors' thermal diodes. In considering different approaches, we have run into the same problem over and over: how do we measure the total heat output of the processor in any nearly accurate way.

The problem with thermal diodes is that between different architectures, there isn't really a way to compare what is being reported. The diodes only measure the temperature at a specific location, and the number we can pull out of a processor isn't necessarily representative of the average temperature of its surface.

Measuring the temperature increase of the system itself has issues as well. Different components in a computer heat up as well, and there's no way to really isolate the contribution of the processor itself to this process. The only number that can truly be compared between processors is the total energy output by the surface of the chip (or heat spreader).

The heat transfer from the CPU to the outside world would be a wonderful number to have, but there is the unfortunate necessity of a heatsink and fan for modern processors to run without cooking themselves. There isn't really a way for us to get to the surface of a chip in order to measure anything. So how are we supposed to measure how hot something is getting when we are required to be cooling it at the same time?

Well, there isn't any easy way we can think of, but just because something is impossible doesn't mean the AnandTech staff can't get it done.
Heatsink Modding: The New Rage


View All Comments

  • PrinceGaz - Friday, April 16, 2004 - link

    Interesting article, but I've a couple of suggestions :)

    If the fan on the heatsink was plugged into the motherboard used to test the CPU, you should instead run it off a dedicated stable regulated power source given an accurate known voltage. That totally isolates any effect the motherhoard, PSU load or anything else might have on HSF efficiency.

    All current and future CPUs we're interested in have a heatspreader, which presumably does a decent job of spreading the heat over a fairly wide surface area compared to core-size of the processor (the AthlonXP/Duron line is effectively finished and won't concern enthusiasts, so their energy-consumption isn't too important). It seems a good way to translate recorded temperature over ambient temperature, into an actual power consumption in watts would be to make a little heater unit (with something like a heatspreader) you can mount the HSF on.

    By adjusting the voltage applied to it and measuring the current you could take a whole range of readings covering a wide range of watts. You'd not only know whether the temperature difference rises linear with power, but also be able to give actual wattage figures for the CPUs. At least one site (whose name evades me) which does extensive HSF testing actually uses such a heater type system to compare HSF performance to ensure identical conditions.

    From the results I must say I'm impressed with how well the Athlon 64 CPUs did, especially when running Prime95. Was Cool 'n Quiet enabled to reduce their speed in idle mode? If it was, I assume you did ensure with some utility they were operating at full speed while Prime95 was running? ;)

    Prime95 is a great way to heat up processors but it may have a different impact on different CPU designs. The P4 runs it *FAR* faster than the AthlonXP does, or even the Athlon 64 despite the latter also having SSE2 instructions (which Prime95 uses). In particular have a look at the benchmarks page:

    and you'll see that for equivalent tests, the time taken by some different CPUs were:

    Athlon XP 3200+ - 0.032, 0.066, 0.142
    Athlon 64 3400+ - 0.025, 0.052, 0.116
    Northwood 3.2GHz - 0.013, 0.026, 0.056

    The Northwood 3.2GHz is running Prime95 roughly twice as fast as even the Athlon 64 3400+. Now while I can't say for sure that running it faster translates into more power-consumption, the fact that the test runs so much faster on that design makes the results difficult to trust. If however you've tried a whole bunch of different stress-test programs on the Athlon 64, Northwood and Prescott and found Prime95 gave the hottest temperatures on them all, thats fine. But if different tests give the hottest temperature for each design, then thats the one you should use for that design when carrying out future tests.

    Sorry for rambling on so long, hope some of that made sense :)
  • CRAMITPAL - Friday, April 16, 2004 - link Reply
  • CRAMITPAL - Friday, April 16, 2004 - link

    Obviously no one with a clue would buy a PressCrap when it clearly runs much hotter than the Northwood and is actually SLOWER in system performance. In InHell marketing speak that's called "90 nano performance". Sic !!! Now inHell is doing damage control after lower than expected profits and defective products they can't even deliver in quantity to Dull and WasteGate. Reply
  • Mumrik - Friday, April 16, 2004 - link

    I meant an articles that claim the AXP only has a difference of 1 Reply
  • Mumrik - Friday, April 16, 2004 - link

    Doesn't seem very well done....

    How can we trust results that claim the AXP2500+ only have an idle/load temperature difference of 1 degree celcius?!
  • Pumpkinierre - Friday, April 16, 2004 - link

    You didnt include the ambient air temp close to the fan (and fan speed- maximum or silent). Northwood temp range seems high. Cpu case temperature can be 10C under internal/diode temp.. 5C or so difference, both idle and load, between N'wood and Prescott is quite significant to an overclocker. The die size is approx. the same but given the 'missing 30 million transistors', 90nm and the large cache, the heat might be more localised on the Prescott affecting headroom. Good to see the a64 data. These AMD cpus seem to run cool.
  • Voodoo80 - Friday, April 16, 2004 - link

    What about cool and quiet? It's a pity this is not tested. I want to know what the idle temps would be if this is enabled. Reply
  • MAME - Friday, April 16, 2004 - link

    Awww, I was hoping it'd be a heat nightmare and AMD would come to save the day! Reply

Log in

Don't have an account? Sign up now