[TriEmbed] Personal supercomputing

[Charles West]

Hello,

I'm looking into machine learning and it seems like some of the methods
could potentially just keep getting better the more data/computational
power you throw at them.  I'm not really skilled enough to do to much with
them yet, but I wanted to go ahead and see what sort of setup it might be
good to build once I have sufficient experience in the area.

Wikipedia has a list of the most power/price computers in the world (
http://en.wikipedia.org/wiki/Flops).  It is interesting to note that the
last two entries are made from commodity PC parts (the latest coming in at
$902.57 and delivering 11.5 TFLOPS).  It would seem that one way to go
would be just build one of these servers with a top of the line GPU.

The complicating factor is that the GPU is really power hungry and takes
something like > .5 kilowatts to keep running.  At normal utility rates,
this means that power is going to cost more than the system if it is kept
operating continuously for more than a year.

The other possible alternatives are building large clusters of Odroid C1s
or Raspberry Pi 2.0s, each of which have quadcore arm processors and only
take ~2.5 watts to run (equivalent power at 200 units).  At the same time,
you could probably only have about 18 units at equivalent cost not counting
energy.

Lastly, you could just build a decked out CPU server.  I don't really know
how they clock in in terms of power efficiency.

A few questions:

What would you build if you had something that would happily eat as many
parallel flops as you could deliver (with correspondingly increasing
performance)?

Does anyone know how GPUs compare to CPUs in terms of power consumption per
FLOP?

At what point does the power cost dominate the computer cost (timescale,
hours of expected operation, etc)?

Also, should this be our new standard way to heat the house during the
winter?

Thanks,
Charlie West
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.triembed.org/pipermail/triembed_triembed.org/attachments/20150205/6e38e899/attachment.htm>