[Beowulf] Re: overclocking with liquids

Robert G. Brown rgb at phy.duke.edu
Sat Sep 22 07:53:49 EDT 2007


On Sat, 22 Sep 2007, richard.walsh at comcast.net wrote:

> Jim Lux wrote:
>
>> The disadvantage of oil?  It's a mess if you have to remove the stuff.
>
> Why doesn't anyone ever mention higher heat capacity, relatively inert gases?
> How does the heat capacity of pure CO2 or N2 compare with air?  There must
> be other candidates that would give you 2 maybe 3 times the heat transfer ability
> without being a mess.  Kind of like global warming ... ;-) ... but inside your
> computer.
>
> My friends at 3M must have thought about this ... maybe, I'll ask.

No need.  Talk to any energy star rated replacement window salesperson
and they'll tell you that the more massive inert gases are better
insulators than the the lighter ones because at any given temperature
they move more slowly.  In fact from the equipartition theorem and a
simple kinetic model:

   3/2 kT = 3/2 m|v|^2

or

   |v| = \sqrt{kT/m}

The heat diffusion rate depends on the mean speed of the particles --
basically they have to move the hot reservoir and the cold reservoir to
transfer heat. Hence argon, krypton, argon-krypton, SF_6, or even CO_2
filled panes (like the ones that surround me as I type this) are good as
they are all molecules that are more massive than O_2 or N_2.  Good for
windows, that is.

Of course, insulating is NOT what you want to do, and besides, diffusive
(conductive) cooling is not what goes on between the surrounding fluid
and a CPU.  CPUs are cooled primarily by convection (well yeah,
conduction to the heat sink and base which are in turn cooled by
convection).  Convection is an active transport of heat, not a passive
one, which is one reason the simplistic argument above breaks down, but
to use convection within a sealed environment capable of trapping a
nobel gas to carry heat out to a chiller strikes me as wrong SO many
ways.  Also note that conductivity is really complicated -- to quote the
wikipedia article on same "There are no simple, correct expressions for
thermal conductivity".  So one is stuck with empirically hunting for
"good" fluids to use for the convective/active cooling process, where
one is trying to beat air in ways that are neither dangerous not
insanely costly.  Which is actually not terribly easy as a COST/BENEFIT
problem (not a physics problem -- the physics is irrelevant where the
cost/benefit is not).

I'd also like to point out that this is a FAT (frequently argued
topic:-) on this list and that there are some great points on it made
(primarily by Jim Lux IIRC) in the archives.  We've kicked around wet
and dry and active and passive and peltier and water.  Under the impetus
of the last such discussion I looked and found that yes, people do make
e.g. water cooled heat sinks (so you recirculate water or ethelyne
glycol or the like in a closed system that is directly cooled elsewhere
to transport heat out of a system).

The upshot of these discussion is usually "You'd have to be mad to
implement wet cooling" in all but a handful of cases, and those cases
generally involve either a very serious raised floor server room where
e.g. water cooling is a cost effective way to achieve some critical
power dissipation density or hobbyists (who are, after all, probably
mad).  The sweet spot of cluster design is one that doesn't make weird
detours into non-COTS regimes, and basic mass market PCs, motherboards,
CPU coolers -- is uses fans, heat sinks, and room AC to regulate the
ambient temperature of the air fed to clusters and transport the heat
back to the AC for rejection elsewhere.  Ultimately you cannot avoid the
basic costs imposed by thermodynamics in terms of P_consumed = P_removed
at whatever operational/ambient temperature you maintain for the cluster
and the ambient temperature of the environment you reject the heat into.

In between -- sure, maybe you can argue that immersing the entire
cluster into a bath of EG or oil of some sort and power circulating same
through heat exchance coils running through a vat of perpetually chilled
water is marginally more effective at removing heat and maintaining the
CPUs at a relatively cool temperature, but at what a cost!  To get at a
motherboard you have to reach into goo.  The goo cannot conduct, burn,
produce toxic vapors or be directly carcinogenic on contact with skin
and has to have pretty good thermal properties.  It has to have just the
right viscosity -- too little and it doesn't "touch" the metal enough to
cool it, too much and it forms an insulating layer on the surfaces of
the metal that is "frozen" out of laminar flow around (like oil, for
example).  Water is actually ideal (non-carcinogenic and all that:-),
except for the conducting part -- even though "pure water" is a
mostly-insulator, it is also a polar molecule and corrosive as hell and
will grab molecules out of its surroundings if it can and transform into
a conductor.

It just isn't worth it, except for the handful of starry eyed mad
visionaries for whom maybe it is.  Me -- I stick with nice, dry, boring
AC at least up to the point where I have to do one of those high power
density racks where liquid cooling a) exists in a OTS form, at least
sort of; and b) it makes economic sense relative to the more mundane
solution of raised floor delivery of chilled air from underneath the
floor that flows up through the rack and is power recirculated from the
top back through the chiller, which is a pretty well understood and
common solution.  Things that are done a lot tend to be cheap, and water
at best is messy.  Non-water is messy and often actively dangerous in
some way.

    rgb

>
> rbw
>
>

-- 
Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu


_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf

!DSPAM:46f5053c156067095717635!



More information about the Beowulf mailing list