Slashdot Mirror
Surface Mapping Athlons For Fun And Knowledge
ryemax writes: "'Surface Mapped Athlon - Is Lapping Required?' is the title of today's bit of insight from FrostyTech. Basically we took an Athlon and measured to within 0.0005" how out-of-whack the surface of the aluminum plate is. With that knowledge and a few hundred values we made a nifty image map of the surface features. With that done, the picture was overlaid atop an image of the processor so all can see where surface flatness may be a problem. Why, you ask? Because I get really annoyed when expensive heatsinks/cold plates get suck on unflat surfaces - and thermal performance gets kicked in the arse. So, rather then just say "it's unflat", I decided to quantitatively prove it using a dial micrometer. Bottom line -- lap that Althon plate." Wow.
but having quite a bit of experience dealing with extreme tolerances (Fixture Builder/Machinists) for 6 years I can say with confidence it really wont matter with this. if you can imagine the heatsink from a horizontal perspective (and if i read that diagram right.. ) it would teeter off on either side about 6/10ths of a thou (just about 1/2 a thou.) Not a big deal, but to put some perspective on the situation, if you consider a dowel pin is undersized half a thou, it should be press-fit (push it with fingers),,. if its over half a thou, ya might have to force it in with a hammer of some sort, nice and tight.. this is apples and oranges though, dowels are cylindrical and this is two horizontal planes.. (did lots of this ugly work with key slot machining) problem is, it costs more and more money to get the perfection these guys are asking for. Calibrating the machines, paying the operater enough so he doesn't slack off for 90% of the day (cus i know it doesn't pay very well for operators),fix fix fix, replace, lots of external costs... and besides, even with it out of tolerance half a thou, i believe its still reasonable for plate to plate heat transfer... those big red areas on their diagram trick the casual reader into thinking *BAD* when it really isn't.. a piece of paper is about 3 thou(sanths) thick, if you take the alumunim foil from a marlboro pack and burn off the one side that has the thin paper on it, thats one thousanths/inch (ya, i machining can be boring sometimes :) and anything under one thousands/inch adds 5X as much money to build :-) Trust me on that. at an old tool&die shop that was working at, some young brilliant know-it-all mech. engineer fresh out of university did some design work for the company.. First project he designed was a battery holder for some Nissan vehicle (i think, something like that, maybe ford)... of course he had to make sure the die was within a .00005 tolerance... ya, a battery holder.. shows ya how much college does , no replacement for experience.
The key issue is to make sure you have a *no* air between the surfaces. Air and other gases have extremely poor thermal conductivity compared with solids and liquids. Therefore air acts as a insulator (the same principle behind styrofoam and sleeping bags).
)
If there is any air between the surfaces, it should be in as thin a film as possible. Better yet, replace the air with a thermal paste. The paste doesn't have as high a thermal conductivity as metal or ceramic, but it will be much better than air.
Thinner films of thermal paste are better, thus (I assume) the concern about flat surfaces. But if I had to pick, I would take a thick film of thermal paste over a thin film of air.
If you want to do a thought experiment on this, the dimensions of thermal conductivity are
Energy/(Time*(Length^2)*(TempDifference/Length)
i.e. BTU/(hour*ft*DegF)
Using these units, thermal conductivity is approximately:
Vacuum = 0.000...
Air = 0.014
Water = 0.36
Steel = 26
Aluminum = 117
Copper = 224
(Source: Perry Chemical Engineering Handbook)
Thermal conductivity does not take convection into account, but convection can be ignored in thin films.
Ok, AMD can make a chip accurate to 0.18 microns. Granted, that accuracy doesn't apply to the top surface of the chip, but the packaging tolerances probably aren't very loose, and I have a hard time believing that Joe Overclocker and a Brillo pad can make matters any better. And how about the heat sink? Is that chunk of aluminum squirted out of a Taiwanese extruding machine really gonna be any flatter than the Athlon?
:)
Whatever.
---
1) The stated value of measuring accuracy is bogus. For one thing, if he did it with a micrometer, you are getting thickness of the plate, not flatness of the plate. It could be totally flat on the heat sink side, and unflat on the other. Plus stupid things like if his body heat raised the temperature of the plate, I bet that whacked off the last measurments from the first by a good .001' atleast, etc., etc. If you are going to say something like .0005' accuracy or whatever it was, put error bars on it, or just don't say it. Misleading.
2) After lapping the processor, then letting it run to operating temperatures, I bet the processor warps some due to differential coefficients of expansion and just cranking the processor on to the motherboard, and the heat sink on top of that.
3) I _assume_ that the the little numbers next to their psuedocolor matrix is a color bar, which shows all of about 1.2 mm of difference in surface relief. When you take in to account the crudity of their measurments, the error bars are probably about that big.
4) Do the math. Heat flux is inversely propotional to the distance between the two surfaces and proportional to the thermal conductivity of the medium and the temperature gradient (q = -(k dT)/l). If you change the distance between the heat sink and processor only a little tiny bit, you only change the heat flow a little tiny bit. It's linear. As long as you increase k by putting heat sink compound, and the distance l isn't huge, then dT mostly takes over.
In other words: Use a good heat sink that stays cool and will make a large dt. Use a thin coat of thermal compound. If your processor still gets too hot, turn down the clock speed a little! Geez. Talk about too much time on your hands.
If I had no sense of humor, I would long ago have committed suicide. -Ghandi
I will forgive you the inaccuracy of the dial micrometer. But I can't forgive the fact that the flatness of a surface can NOT be measured with a micrometer. You need a known flat surface (certified granite plate), and a dial indicator with sturdy, stable stand. Anyone who has worked in a machine shop can verify this for you. The way that you arrived at your measurement; the surface possibly, could be "perfectly" flat.
Standard-issue thermal grease has a higher specific heat than the metals used on the heatspreader and the heatsink, so less heat is transferred between the processor and the heatsink. More/closer metal-to-metal contact is always a benefit.
kugano
Yes, the flattening effect is nullified unless you also lap the bottom of the heatsink. Yes, a little thermal paste accomplishes almost (but not quite) the same thing. BTW, most people put about 10x the grease on there that's actually necessary. The layer should be about as thick as a sheet of notebook paper. Much more and you're not helping, you're hurting.
The main benefits of lapping come from two points unrelated to the flatness of the contacting surfaces. When heat's conducted through different media, each change in medium reduces the overall conductivity. Some processors (i.e. Celeron) have jacket over the casing. The jacket and the casing are two different materials. If you sand the jacket away, that's one less medium to conduct through.
The other main benefit is from reducing the thickness of the casing between the processor core and the heatsink. If you're dumb enough to sand right through the core, you were too dumb to be taking sandpaper to your processor anyway.
Yes, I overclock. No, I'm not ashamed of it. I overclock a Win9x box used for gaming and for visitors who get freaked out when they can't find a Start menu. Since the stability of Win9x approaches zero, knocking it down a few percent farther produces generally no observable effect.
Now, you're all more than welcome to flame me (and what kind of place would Slashdot be if you weren't?). Before you do: Yes, I'm aware that overclocking and lapping generally shouldn't be done in an environment where stability is important. Realize that some of us prefer a side order of hardware hacking with our RDA of software hacking.
Note to moderators of this discussion: Posting something like
'Doing this would be fscking stupid. I don't really know anything about this, and I've never tried it.'
is called trolling.
J
Happy overclocking! I've got my Athlon "700" running at 900, and it maintains 30c!
A bit of advice - always spring for a 300W+ power supply with Athlons...400W if you plan to use the Peltier!
"History doesn't repeat itself, but it does rhyme." Mark Twain
As the human head radiates approximately 75% of all our expended body heat, we could use the lessons learned from this site to efficiently cool our pointy noggins, so as to allow us to overclock our craniums to the maximum extent.
At the very least, cranial cooling will reduce the number of times we product erronious errors, or outright crash.
Damn, I forgot to take my medication again.....
Feed The Need[goatse.cx]
If it's flat when it's cold, is it flat when it's hot?
.001 inches.
The coefficient of linear thermal expansion is:
Aluminum: 25 x 10-6 (C)-1
Silicon: 3 x 10-6 (C)-1
Difference: 22 x 10-6 (C)-1
If you attach a sheet of aluminum to a wafer of silicon the aluminum will expand more than the silicon. The difference in the expansion will be 22 x 10-6 of an inch per inch per degree centigrade.
For a 4 inch piece of aluminum heated ten degrees the difference will be approximately
If the plate is fixed at the ends it may bow out. The amount that the bowing will pull the aluminum away from the silicon is approximately:
B=Square root of ((L+sigma)2 -(L)2)
Where L is the length and sigma is the expansion difference. For sigma much smaller than L this is approximately:
B=Square root of (2L*sigma)
If L=4 inches and sigma=.001inches
B=Square root of (.008)
B=.09 inches
That's almost a tenth of an inch. That's on the order of fifty times larger than any surface imperfections.
Now, I've admittedly made a lot of simplifying assumptions in my calculations. Some of them, if anything, underestimate the bowing factor.(Only ten degrees above room temperature--that's a well cooled processor indeed.)
But the magnitude of the effect of any bowing due to differential thermal expansion is so large, that if it does occur it would dwarf other departures from flatness.
So attention must be paid to how and where the alumium is attached to the silicon (as well as joints between other substances) to be sure such thermal bowing does not occur when the computer is actually running.
The Eponymous Mallard
If it walks like a duck...it's the Eponymous Mallard
they tried various options with good results. They had the computer running in the freezer
Towards the end, things got hazy. To quote:
some very interesting photos too
"It is a greater offense to steal men's labor, than their clothes"