Add your Beowulf cluster here

Conundrum, Mon Mar 21 2016, 05:58PM

Link2

Interestingly the price point for the P940 AMD CPU is low enough that any old Sony Vaio PCG-61611M or similar motherboard (of which there are lots to be had surplus) can be leveraged into a cluster using 38V at 10.7A

My target is 32 1.7GHz cores, 2*2GB per node (negotiable, about the minimum is 2*1GB DDR3-10600) and 8-16GB Flash per node with the option of networking the nodes via TOSLink and adding load shedding to keep the power usage within safe limits.
This is for eight laptop boards assuming modified BIOS optimized for maximum speed and running the fans at between 30 and 125% speed when doing actual computations.

So far I've sourced one working board, one DOA-but-probably-stuffed-BIOS and another three which are <£15 each.
Can weed out the really bad ones using a simple voltmeter test and harvest bits from these, turns out that BIOS and LCD problems don't affect the VGA output as its actually the EDID clock line which is typically stuffed.
EDIT: And a small fuse near the VGA port also goes bad, as does the inverter, LCD screen and CPU voltage control IC.

My earlier attempt used Atom based machines but I had that one up to 5 working boards before it became obvious that even overclocked the system wasn't up to the job.
A single quad Core 4 would have been better here and have cost only about £300 at the time, compared to £60 but most of this got recycled and sold on installed in refurbished AOA150s.
also see Link2 Very cheap, thermal loading might be an issue but this is fixable with the method described above.
Re: Add your Beowulf cluster here
hen918, Mon Mar 21 2016, 06:31PM

I would love to do this, but for two things:
The sheer inefficiency
The fact that modern graphics cards beat any CPUs for parallel processing. My Radeon R9 290X can deliver 5,632 GFLOPS, the desktop version of that CPU can deliver 40 GFLOPS

By my calculation, you would need 141 nodes consuming a stupid amount of power (7166W) whilst I can sit here drawing 350W with CPU and GPU on full load. You would beat me in GBs of RAM, though it wouldn't be very fast RAM.
Re: Add your Beowulf cluster here
Conundrum, Tue Mar 22 2016, 06:58AM

Yeah, this is indeed the case.
£90 gets you a used Nvidia Tesla good for 1.2 TFlops if used with optimized thermal path and finest silver nanoparticle heatsink grease.
See Link2
Good luck getting more than one of these to work in all but the most recent MB though, they will typically work but be hobbled by the x16 bandwidth of a cheap board which shares one southbridge chip between 4 (usually adjacent) PCIe ports.
For a mobile robotics system the miniature Nvidia Jetson TX-1 boards currently in production as of March 2016 will yield PS4-like performance (1Tflop) drawing under 30W of power when running maxed out.

EDIT: Looking at crowdsourcing some sort of small supercomputer for the hackerspace.
Because having our very own supercomputer would be seriously cool even if it was only 1/1000 the size of Tianhe-2.
Just for perspective, Lt Cmdr. Data of ST:TNG fame had a peak performance of 60 Tflops per second.