Slashdot Mirror
Extreme CPU Cooling
Darkfell writes "Check out todays HardOCP.
An article was posted by a guy who cooled a
dual PIII down to -59.7C. Very nice setup." This
is worth a read- quite detailed for you do-it-yourselfers
willing to risk destroying your computer.
We've done it using dry ice. The chip dropped temp to around -100C. It ran perfectly. We managed to take a P120/60MHz bus => P200/100Mhz bus. Stable.
It can be done. We've done it. Thermal contraction is only an issue if you cool it *too quickly* - we're looking into nitrogen for the next test. Passive submersion - the whole board. Should be interesting. But to answer your question again - thermal contraction until about -100C is a joke - don't worry about it. Much.
--
He'll run into problems when he tries to reach -80, though. Sooner or later traces in the chips or on the chip modules will crack due to differing rates of thermal expansion in the materials used. An interesting read nonetheless, though.
Yes. In fact, they operate more efficiently, which is why he could boost the clock rate. This is also why processors fail at higher temperatures; they work _less_ well as the temperature increases. If I understand correctly the transistor threshold voltage and a few other parameters vary with temperature. I'd have to dig out my old electronics textbook to give you a detailed explanation, but the gist of it is that the transistors end up passing more current, which decreases switching time.
Limits to clocking with this kind of scheme are chip failure due to electromigration (the traces in the chip can only take so much current), and chip or module failure due to cracking caused by different rates of thermal expansion in the materials used.
This doesn't address the problems that I raised - thermal expansion/contraction difficulties and electromigration. While you do mention the limits imposed by the way the threshold voltage changes with temperature, the other two factors mentioned may be what limits your ability to cool and overclock chips.
An integrated circuit chip is a chunk of silicon with aluminum wires on top of it embedded in a thick layer of silicon dioxide. These three materials have different coefficients of thermal expansion. As you cool them, they will change sizes at different rates, causing stress in the chip. Cool them enough, and your chip will break. I'm told that the temperature at which this occurs is lower that I had originally assumed, but it *will* happen. Possibly with liquid nitrogen (though some successful liquid-nitrogen-cooled systems have been built), and almost certainly if you do something silly like cool a chip with liquid helium. Likewise, the card on which the chip sits is glass fibers in an organic resin with plates and traces of copper. These materials all expand at different rates. You also have a lead alloy connecting the pins of the chip to the copper traces on the board. Size changes due to temperature will put stress on these solder joints - and size changes in the plastic casing that holds the pins and integrated circuit chip will put a lot of shear stress on these weakened solder connections.
So, this is not something that can safely be ignored forever.
Likewise, electromigration will seriously reduce the lifetime of any chip being run significantly faster than its standard clock speed at room temperature. Electromigration is the tendency of metal atoms in wire traces on a chip to flow along with the direction of current. The higher the level of current, the greater this effect. If a chip is operated with too-high current levels for too long, enough metal atoms flow that the trace becomes brittle enough to snap under mechanical and thermal stresses, or develops a gap, or gets thin enough that resistive heating melts the trace or increases its resistance enough that it can't transmit signals properly. Electromigration effects were a common cause of failure in older chips. To compensate for this, chip designs nowadays specify maximum currents for given sizes of traces and are careful not to exceed them. However, overclocking _does_ exceed them. In order to clock a chip more quickly, you have to charge and discharge the parasitic capacitances within it more quickly - which is done by increasing the amount of current flowing through transistors in their "on" states. Conventionally, this is done by cranking up the core voltage. Cooling does this by lowering the threshold voltage and fiddling with a few other transistor behavior parameters. In both cases, for a chip clocked n times more quickly, you have n times the amount of current flowing through the same traces. Clock a chip at twice its normal speed, and you have twice the current through the traces - and a chip that will burn out far sooner due to electromigration. Copper is more resistant to electromigration than aluminum, but the metal traces in copper chips are correspondingly thinner than the metal traces in aluminum chips. This is why copper was adopted; the aluminum traces had to be made wide enough that bulk and parasitic capacitances were becoming real problems. Modern chips - copper or aluminum based - are designed to run just below the threshold for electromigration damage. Overclocking them to the degree being done with these cooled setups _will_ push them over the threshold.
OTOH, if you don't care if the chip burns out in a year or two, go for it. Just be aware of the limits and side effects.
I've been interested in this sort of thing for a long while and have been overclocking since the old 8088 days when it took a soldering iron and crystals from the local 'Shack.
Anyway, cooling with refrigerant has never really appealed to me much. There may be some danger in working with the Freon, I'm not terribly familiar with the technology involved, and the power drain could wind up being significant. Some of you have posted about simply placing a computer in a small refrigerator - this has already been done and it was found that the compressor in the 'frig couldn't keep up with the heavy heat load of an overclocked CPU - the compressor ran full time. The OCP article mentions this problem too, frankly I found the article very well written covering many bases most people forget.
Anyway, my focus has been to build a water cooled Peltier assembly. I currently use Peltiers to cool a PPGA 300A enough to go 504mhz (stable) but the heatsink become significantly warmer due to the Pelt's heat. Obviously such a setup isn't ever likely to go below ambient either.
Some have mentioned condensation and water as being really big issues. So long as the water is sealed out of your sink this isn't an issue. Condensation can be avoided by decent insulation - remember that condensation only forms when components cooler than the dew point meet humid air. Avoid this and you're fine.
One of the last hurdles to consider (IMO) FSB speeds. Currently there just aren't enough selections and Intel's damned multiplier locking is making life a bitch. The Turbo.PLL the Japanese are working on may fix this as it'll allow you to vary the FSB in increments while keeping things liek the AGP and PCI cards at a normal speed.
Way below are some URLs to check out. Note that some are in Kanji as the Japanese have really had a good time with this. Note too that Melcor sells components to water cool Peltiers for industrial applications and apparently not retail, someone needs to resell these parts! Lastly, the Socket 7 CPUs and the PPGA Celerons share a common size, the PPGA chips also appear to run cooler than their slot one brothers. My fastest systems all run the PPGA Celerons including one dual SMP system that's not actively cooled but still gets 2X464mhz. Note that I've not yet managed to get a successful water system running but am working on it. Car heater cores work well for heat exchanges, RedLine Water Wetter helps improve heat transfer, and small fountain pumps move massive amounts of water - these are designed to be constant duty too. All Electronics sells Pelts cheap BTW.
On with the URLs! Here are just a few of what I've got and I'd welcome correspondance on this subject if my HotMail 'box can handle it!
------------
http://www.melcor.com/ - Industrial hardware cooling supplier
http://www.agaweb.com/coolcpu/ English water cooling site w/plans
http://e-sdi.com/west/intro.htm English water cooling project, self contained
http://www.mune.com/mcp2.htm Japanese site, Kanji w/Multiple projects shown.
http://www.kumagaya.or.jp/~touma/index.html Japanese hardware site - Kanji
http://www.jah.ne.jp/~ken1/kenO.htm Japanese project - Kanji
I'd post more but after 5 Netscape crashes I've got to run. Explore the Japanese sites and check out the Turbo.PLL if you happen across it - that site is slowly being translated. If this is of real interest I'll try to post more URLs when I've more time!
Enjoy!
P.S. no time to preview, hope it comes across okay!
Build it, Drive it, Improve it! Hybridz.org