Slashdot Mirror
The Amazing Shrinking Supercomputer
mE123 writes "It would seem that IBM is trying to change what we all think of as super computers. Their new Blue Gene family of super computers is meant to be 6 times faster, consume 1/15 of the power and be 1/10 the size of current models. The prototype is already number 73 (with 2 teraflops) on the list of the most powerful super computers and it's only "roughly the size of a 30-inch television". They are hoping to be able to make it up to 360 Teraflops using only 64 racks." We covered this a bit earlier, but without the level of details.
No exact dimensions, but there are some photos here.
Who do you get to be an expert to tell you something's not obvious? The least insightful person you can find? -J Roberts
Cray made the Cray EL series from '94-'97, they were a "deskside" computer. See here for more info.
Trolling is a art,
Space and heat dissipation are becoming very serious limiting factors in the scalability of supercomputing clusters.
In theory, you could just keep adding more and more nodes to an existing system, and as long as your interconnects were good enough, you could scale.
But in practice energy consumption (and getting rid of the waste heat afterwards) will hit you before you can get much futher that we are today. The Big Mac G5 cluster in VA, for example, required custom cooling systems because conventional aircon units simply couldn't handle the load.
As a result, IBM's work is *vital* for making faster supercomputers -- and the improvements they're claiming are very impressive indeed.
You can find most of what you want to know on IBM Research or US Department of Energy (search for bluegene). I think both can survive slashdotting.
Top500.org
Top 500 Supercomputer list
Where law ends, tyranny begins -- William Pitt
It is much more difficult to use them for most applications most of us can think of. For example, VLSI CAD software (simulation/analysis/synthesis) is very compute intesive. However, these systems usually do not even take advantage of the multiple CPUs in a typical general purpose SMP system. You have to manually partition designs and sometimes loose the advantages of global optimization.
So don't run and order your new Blue Gene yet :)
Why do we need to have small, power-efficient supercomputers? Isn't the main goal of the supercomputer to be fast as hell? Granted, if this can be achieved while simultaneously minimizing power and size then by all means go for it. However, as stated by my parent, what sacrafices are being made?
The increase in speed is related to the reduction in size.
For a moment, let's pretend that electricity within a wire travels at the speed of light.
Now, let's pretend that we wish to carry pulses of electricity from one end of the computer to the other at a very high speed.
At some point, the distance the signal has to travel will become significant to the speed of the computer.
This is already happening in PCs. If you take a close look at the motherboard in your computer, chances are you'll see weird places where the traces just zig-zag back and forth (notice the angles on them, that's not by accident either, but I'm not going to try to explain a fourth-year university course in microwave and RF design here). These zig-zags add length to the traces so that they have the same length as other traces within the same bus, and all the signals on that bus arrive at the same time. Think of them as being "equal length headers", if you're into the throb of a big-block V8.
Length of interconnecting wires is non-trivial at this point. Stray capacitance and inductance caused by any conductor are non-trivial at this point. As a result, a terrific limiting factor to the speed of a computer is now its size.
Power consumption is also related. Modern ICs are made of millions of MOSFET transistors which behave as switches. These switches are not perfect: during the transition between a logic high and a logic low, the transistors spend time in the linear state where they are resistive. As a result, they waste energy as heat.
Stray capacitance and inductance - even within the junctions of the transistors themselves - slow their ability to switch instantaneously. As a result, they must be made as small as possible to reduce capacitance (C) and inductance (L).
This also explains why newer generations of a processor can run faster than their predecessors: smaller and smaller features on the IC mean less stray C and L, which means that the transistors can switch states faster, which means that they spend less time in the linear state and therefore heat up less. This means less energy wasted as heat.
Fire and Meat. Yummy.
According to this rather detailed paper, IBM designed the unit around 700 MHz (0.13 micron) PowerPC 440 Processors. This is not the modern equivalent of a Cray 3.
The node to node density, though, is very high. The maximum cable length is 8m.
This set of sldes compares some of the architecture of the BlueGene/L to other ASCI machines.
They once made a machine out of FPGA's. It worked by evolution: It would rearange different FPGA's and work out which gave the correct answer the quickest, and learn from there. Basically, it was pretty slow the first time it tried something. But if you let it learn for a while, you got supercomputer performances out of a tower-sized box (On the specific set of tasks it has learned, anyways).
Its good for plenty of fixed-task things: Medical imaging, software-defined DSP, scientific computing, that sort of thing. You can check them out here
ASCII stupid question, get a stupid ANSI
Seems to me there was once a guy named Seymour, who could do the fast *and* small thing quite well. Since we're talking in terms of Televisions...
Big screen #1
40" HDTV and a A size perspective
We have DEFINITELY been down this road before folks. I don't see why it's so hard to do this, unless you're using COTS components. Hence, the point of "engineering" - not cramming a bunch of stuff in boxes/packages into bigger boxes and packages.
Detachment 3 Media
Exposed, Exploited, Exploded
This one little computer is small and efficent and all the waste heat easily taken care of. Now imagine not just one of these, but a whole building of these. Our heat problem crops right back up.
IBM knows what it is doing.
That being the case, why aren't distributed apps considered as part of the Super Computer list?
Most of the tasks you pick a supercomputer for aren't things you can cut up into a thousand chunks and let every computer finish it's chunk of the problem independently. In particular, the benchmarks (LINPACK) that determine who goes where on that supercomputer list generally measure a computer's performance at big linear algebra problems (which are what takes up most of the compute time for huge classes of real problems), and for those problems every node needs to share results with many other nodes after essentially every iteration: this means you need high bandwidth and very low latency connecting the nodes.
Now, the supercomputer benchmarks may make things worse than they have to be: according to this they're measuring performance on dense matrices (where every node needs to talk to every other node), whereas many real world problems can be discretized into very sparse matrices (where each node only has to talk directly to a few of the others) instead - still, even in the sparse situation you want your computers to be separated by microseconds across your high speed interconnect rather than milliseconds across the low bandwidth internet.
I recently did a search of top500.org which has specs back to June 1993 and up to June 2003
BTW WHO THE HELL BROKE top500.org!?!? This site used to be easy to use and informative, now it is a banner add hell, that obscures the info you used to be able to get to easily, with many broken links and apologies for works in progress.
Anyway I digress, the point is that in 1993 the fastest computer was the TMC at Los Alamos with GigaFlops ratings of 59.7 Rmax 131.0 Rpeak
My Dell XPS today would rate in the top half of fastest machines in the world for 1993 if I'm reading the stats right with just over a GigaFlop of power.
Todays fastest machine is Japan's Earth Simulator rated at 35860 Rmax 40960 Rpeak
If we define a super computer as the ability to get in the top500 then 245.1 Rmax 384.0 are the numbers that indicate your machine would be a super computer by 2003 standards.
Letter To Iran