Reasons to not buy a CPU

[Floid]

Yep, Athlons have been warm.  The thing people forget is, they're rated to *take* the heat they produce.  As long as it's kept within the range specified in the datasheets, you can expect it to tick away as long as anything else, or until your Taiwanese capacitors explode.

Meanwhile, if you've missed the chatter on the brit sites (Inq/Reg), they're doing something ridiculously smart.  Everything in the Hammer (AMD64) line runs cold now, but they won't specify the current dissipations - only the *peak* they'll ever allow for the model line, capped at a toasty 89W.

Wait, isn't that ridiculous?
http://theinquirer.net/?article=11948

Aha!  Any "Opteron/A64-ready" heatsink ever made actually *will* be ready to take the load a year or more down the line.
http://theinquirer.net/?article=11964 - Third letter down.

Future-proofing from an x86 vendor.  Who'd've thunk.

[Floid]

mikeymike wrote:

1 - electrical conductivity improves if components are cold.  Equals faster.

BIG pet peeve.  "Faster," yes; but the 'fastness' is the switching speed the transistors can reach without frying.  (More technically, transistors require a certain amount of current, interconnects have to handle a certain amount of current, get things too warm and resistance goes up, interconnects toast, and you're out a chip.)  In other words, the clock rate.

Cooling lets you, or the manufacturer, bump the clock rate.  ("Overclocking," when it's done by a user without industrial-grade testing and QA.)  It won't make your 2GHz faster than anyone else's.

2 - if you switch the computer off at night, all the components cool down, and with today's machines that can mean a reduction in temp by about 20 degrees C.  Heating up and cooling down something repetitively is a great way to damage it, killing it far more dramatically than if it was on all the time.

Quite true.  We shouldn't forget that top-tier vendors have and do throw a lot of money at these issues; I recall AMD had a batch of Athlons where the BGA? of the core didn't register properly on the package (pop, fizzle, zzt) and Intel's had rumors of melting Xeons, but quite frankly, you see a lot more of it from vendors forced to play catch-up.  I've seen a lot of iBooks with heat?-warped touchpads, the original 1000 was a first-run design with its share of little issues like that (okay, grounding, not so much heat), and so on.

Reliability is so important to the top tier that, barring the occasional lemon (end-of-line cores, from either vendor), they've made it pretty hard for people *to* screw up.  What causes the most pain these days are the unpredictables - Taiwanese capacitors, the bad package material that brought Fujitsu down, whatever the heck it was that made the IBM Deathstars so flaky (predictable or not?) - and so on.  If you follow the datasheets and design guidelines, your design will probably make it to the MTBF.  Even without any common sense.

3 - Hard disks die much quicker if they're running hotter.  They may be rated to work at {insert high temperature here}, but the specs don't show you how that effects the MTBF (mean time before failure).  Personally I care far more about the data than most of the other components in my system.

There's a lot of chaos at work there.  To some degree, heat can be good; it keeps the oil thin and slick in your car engine, for instance.  Crank it up too far, and you turn it to gum, or cause some other aspect of the mechanism to fail before its time.  Drives die for so many reasons that it's hard to say what degree of cooling you really want; ventilation surely helps, but running 20 case fans or Peltiers is enough to cause a voltage sag that can screw you up in other ways.

[Floid]

mikeymike wrote:

1 - electrical conductivity improves if components are cold. Equals faster.

BIG pet peeve. "Faster," yes; but the 'fastness' is the switching speed the transistors can reach without frying. (More technically, transistors require a certain amount of current, interconnects have to handle a certain amount of current, get things too warm and resistance goes up, interconnects toast, and you're out a chip.) In other words, the clock rate.

How can they fry if they're cold? I'm not saying that because they're cold that the clock speed can be raised, while that does stand to a certain extent, I'm saying that components that are running very cold make better conductors of electricity.

They're called 'semiconductors' for a reason; they have to be able to create resistance to function.

So if we extrapolate ad absurdum, turning your CPU into a superconductor ain't gonna help.  (Turning the *interconnects* into superconductors might be a bonus, sure; ring me when you've found a way that can be implemented with current processes.)  Now, keeping the temperature within a range may in turn lower resistive losses during the 'on' state, but that's the range they came up with when they binned your processor and wrote the datasheet.  "Improving the flow of electricity" is really not something that concerns the end-user here; if you're freaking out while running the chip at its rated spec, direct the cryogenics to the voltage regulators and the traces that deliver current to your graphics card.

Also note that, while, yes, ultimate cooling might prevent destruction, *point heating* is a big concern.  If you can demonstrate any cooling solution that allows you to shunt wall current through a CPU and have it live to tell about it, again, pick up the phone.  In practice, there's a law of diminishing returns; eventually, the thermal conductance of the chip package and silicon itself become limiting factors.  Addressing this is of major interest to the chip industry - it's nice to be able to keep following Moore's Law without causing an exponential increase in the price/mass/size of cooling solutions to the consumer - but under normal circumstances, it should've been long-since vetted before it got in your hands, and there's not much you're going to be able to do about it afterwards.  (Unless you have a miracle thermal conductor, and the equipment to shave the package  down to apply it - no, Arctic Silver and some sandpaper for lapping doesn't really count, see opinion in the next paragraph.)

In other words, if your chip is hanging on by one transistor that'll die if the temperature rises a degree K (while still being within the rated spec for the chip) ... Cripes man, let it go.  Better the core kicks the bucket than you risk subtle bit errors the moment your fan collects dust and loses an RPM.  I'll make an exception for 'rare' chips (and an exception from 'normal circumstances' when they get to the end of core lifetimes and just start seeing what they can bin up and dump on the market), but really, either everything's going to take well in excess of what it's rated, or you're living every minute near the brink of disaster anyway.

And yes, this is why I bought a Thunderbird at 850, not 1.4GHz or wherever they took it near the end.  (But SOI seems to change the game again; reducing leakage currents seems to imply you hit architectural clock limits long before you hit killer heat.)

whatever the heck it was that made the IBM Deathstars so flaky

I think that must have been well and truly IBM's fault, as they haven't fired off a silo of lawsuits at anyone.  :-)

The world may never know, more's the pity.

To some degree, heat can be good; it keeps the oil thin and slick in your car engine, for instance

PCBs and electronic components are NOT care engines!  They don't share any common characteristics!  Car engines contain 99.9% moving parts.  Computer components have very few.

Yes, and they're all in the HD.

Now, from what I can find, heat is apparently a joykill with fluid dynamic bearings (though what "heat" involves isn't clear), but running them below 273K probably wouldn't be any healthier than pouring liquid N2 down your pants.