Slashdot Mirror
How To Build a Supercomputer In 24 Hours
An anonymous reader writes with a link to this "time lapse video of students and postdocs at the University of Zurich constructing the zBox4 supercomputer. The machine has a theoretical compute capacity of ~1% of the human brain and will be used for simulating the formation of stars, planets and galaxies." That rack has "3,072 2.2GHz Intel Xeon cores and over 12TB of RAM." Also notable: for once, several of the YouTube comments are worth reading for more details on the construction and specs.
Computers have surpassed that level a loooooong time ago
Doubtful.
The computational requirements for simulating the human brain have been severely, even hilariously, underestimated in the past. To quote Wikipedia: One estimate puts the human brain at about 100 billion neurons and 100 trillion synapses.
That's... a lot.
First off, a lot of people think that 1 FLOP = 1 Neuron, which is not even close. The active points are the synapses, of which there are about a thousand per neuron! Each may receive an impulse over ten times a second, and involve dozens of parameters, such as the recent history of firings, neurotransmitter levels, hormone levels, membrane potentials, etc... A very conservative estimate would be that a single neuron, receiving impulses at around 10 Hz on 1000 synapses would require on the order of 1 megaflop to simulate. That's ONE neuron. Now multiply that by 100 billion, and you get a picture of what's required: about 100 petaflops, minimum. Storage is nothing to sneeze at either. Assuming a mere 50 single-precision floating point values per synapse to store all simulation state, you're looking at almost 18 petabytes of memory! That's over $100M for the memory sticks alone, even with a deep bulk-purchase discount. Unlike most server or HPC workloads those 18 petabytes would have to completely read out, processed, and possibly updated again at least ten times a second.
Second, consider that the first simulations won't be very optimized. We still don't really know what's relevant, and what can be simplified away. Hence, I suspect that the first attempts will be much less efficient, and may require 10x or even 100x as much computer power compared to later attempts. For example, neurons don't just fire impulses, they also grow and change shape. I don't think there's even a good model for how that works in the complex 3D environment of the brain!
We are getting closer, but expect to wait at least a decade or two before people start talking seriously about a full human brain simulation.
So these things happen on different (connected) machines if you will. Every neuron is a processor in itself anyway.
I think that's part of why computing power greatly surpassed humans long ago, and will not reach human levels for many years. The brain isn't digital. It holds an "infinite" number of analogue states, simultaneously. With massive errors and gaps filled in with guesses made from other parts, without even an minor error check that indicates that he information being determined to be "true" is 100% interpolation with 0% fact or actual memory. The very idea of an error check that was wrong more than right and kept no indication of where the result actually came from is so incredible that nobody would ever create a computer capable of operating that way. It won't be until we have computers many millions times more powerful where we can remake a "perfect" brain, until then, we'll never be able to match the capabilities of the human brain.
Learn to love Alaska
it took ~year to acquire the funds, benchmark tests, fix the design, make the tender for the parts etc, but all the construction was done in 3x8hr shifts
we use various astrophysics simulation codes, i.e. GASOLINE, PKDGRAV, RAMSES etc. some are developed by us. they are all MPI and solve the coupled gravitational and hydrodynamic equations that can describe the dark matter and baryons evolving in the expanding universe. memory and speed of the computer limit the resolution that can be attained, so various "sub-grid" physical processes have to be treated carefully. for cosmological simulations we know the initial conditions - those are the fluctuations that we can read off the microwave background. they show the universe was hot, dense and smooth early on. the codes follow the perturbations into the non-linear regime when dark matter haloes, stars and galaxies form. we can then compare the properties of simulated structures with observational data etc.
surprisingly few - a couple of bad motherboards (or static ;). its only been up for a week or so and we are still testing/installing stuff before making user queues live.
Used to do this kind of stuff when I was with IBM about 10 years ago, we had a group in XSeries Manufacturing who specialised in quick turnaround configuration of HPC rack systems just like this. Funnily enough, one of the major logistical elements was dunnage, ie the empty cardboard/foam and plastic that all the option parts arrive in. When running full out we used to have 1-2 guys per shift just to move the rubbish out to the big compactors out back. You wouldn't believe just how much packaging even a comparatively small cluster like that can generate.