[Beowulf] liquid cooling for HPC

Jörg Saßmannshausen j.sassmannshausen at ucl.ac.uk
Wed Dec 5 05:40:31 PST 2012


Dear Jim,

the motherboard is really immersed in this fluorinated liquid which acts as a 
heat transfer medium here. This heat is then transferred to the cooling plate 
which is where the water channels are coming in. The whole box is sealed so 
there should be no escape of the fluorinated liquid here. The only possible 
leakage would be the water connectors here. This water is pumped around in a 
closed loop and you need a heat exchanger to remove that heat. Here you could 
use that heat to heat up a cold corridor etc., similar as you do with a heat 
pump in your house.

I hope that clarifies that a bit. Note: all of that is my understanding of the 
system as I saw it at the workshop.

All the best

Jörg

On Wednesday 05 December 2012 12:38:56 you wrote:
> Interesting..
> I can't figure out how they are transferring the heat from mobo/components
> to the fluid. On the one hand, it looks like they're immersing the board
> in a sealed container, but there's also what is identified as "water
> channels" (page 2 of the pdf)
> http://www.iceotope.com/assets/files/pdfs/module-a4-icm-t1-si2-spec.pdf
> 
> 
>  They've got the usual sort of "liquid connector" things which probably
> don't leak too much (and you're much better off than in, say, a hydraulic
> system where inevitably, there's grit and such that gets stuck in the
> valve built into the connector).
> 
> Liquids leak, always, but with decent design in the rack, at least the
> drips don't go somewhere bad.
> 
> Using the "engineered fluids" seems like a typical approach.  Back in the
> day, Fluorinert was popular (but expensive).  It looks like they're using
> conduction for the heat transfer, as opposed to ebullient/condenser
> strategies.  (but if they're using that, why do they reference "water
> connectors and water channels"?)
> 
> Or do they have the board immersed in a suitable fluid filled box, and
> there's a fluid/water heat exchanger built into each little module?  That
> makes reasonable sense.  Seal the expensive oil/fluid in the module, use
> water (with corrosion inhibiters mixed in) to circulate around.
> 
> 
> 
> 
> On 12/5/12 1:50 AM, "Jörg Saßmannshausen" <j.sassmannshausen at ucl.ac.uk>
> 
> wrote:
> >Dear all,
> >
> >I am sure the subject has been covered a number of times, however, I have
> >been
> >to the Machine Evaluation Workshop last week here in Liverpool and this
> >company here
> >http://www.iceotope.com/
> >done something which I found rather unusual. Rather then using pipes to
> >cool
> >the CPUs or dipping the whole node into silicon oil (messy) or water,
> >they are
> >using a hydrofluoro carbon compound to move the heat from the motherboard
> >to a
> >chilling plate. There are pros and cons of course. As the whole
> >motherboard is
> >sealed into the box you cannot change it yourself. It needs to be send
> >back to
> >them. On the flip side, it appears the the excess heat could be used a
> >bit more
> >efficiently.
> >I was wondering whether somebody has some experiences with 'water'
> >cooling in
> >HPC and would like to share it with me? I am thinking of getting
> >something
> >like that or similar (no decision has been made right now) for the next
> >server
> >I am getting. However, as that is new technology to me (the 'water'
> >cooling),
> >before I venture into that I would like to hear some first hand
> >experiences,
> >not necessarily with the company/product mentioned above which is only an
> >example here.
> >
> >All the best from a wet London
> >
> >Jörg

-- 
*************************************************************
Jörg Saßmannshausen
University College London
Department of Chemistry
Gordon Street
London
WC1H 0AJ 

email: j.sassmannshausen at ucl.ac.uk
web: http://sassy.formativ.net

Please avoid sending me Word or PowerPoint attachments.
See http://www.gnu.org/philosophy/no-word-attachments.html



More information about the Beowulf mailing list