Self-wiring supercomputer

Researchers at the Edinburgh (Scotland, not Indiana) Parallel Computer Center are building an energy efficient one-teraflop supercomputer about the size of four PCs (a 2GHz P4 runs at just a few gigaflops). While modest by super-computing standards (hundreds of teraflops), this type of horsepower usually requires 100 times the energy consumption and enough silicon and plastic to fill a room. The power and efficiencies stem from the use of 64 Field Programmable Gate Array (FPGA) chips (instead of conventional microprocessors) which can be "reconfigured using software to mimic computer processing equipment that is physically designed to take on specialized tasks." And therein lies the rub — FPGA hardware is more difficult to program 'cause a programmer must understand how to tweak the underlying hardware. Well, let us know when you sort it out 'cause we're jonesin' to make a hook-up: Mr. Self-Replicating robot meet Mrs. Self-Wiring supercomputer. Then we'll silently slip on our black-and-white Nikes and await the birthing of our synthetic overlords. 

[Eric @ Dec 19th 2005 2:05AM]

Isn't Sony claiming one teraflop for the PS3? I could have sworn I saw that on the e3 video.

[zee @ Dec 19th 2005 2:05AM]

Eric you sound shocked that Sony "lied"

[Wei @ Dec 19th 2005 2:05AM]

The Xbox 360 has 1.something teraflop. The PS3 claims to have 2.0 teraflops...

[Patrick Flynn @ Dec 19th 2005 2:05AM]

Teraflop? What Does That Measure? Spped, Heat, Or Power Consumption?

[gabe @ Dec 19th 2005 2:05AM]

I can tell you from experience that the Xilinx software that controls the FPGAs is a total pain. There could have been advances in the last 3 years that has made it better, but I'd constantly fight the little problems you'd run into that would eat up time, and at 4am those things really bother you...

[Ian Jardine @ Dec 19th 2005 2:05AM]

"I self wire, therefore I am"

[sparc @ Dec 19th 2005 2:05AM]

I'll second the motion that xilinx software sucks. It's some of the clunkiest software I've used, but I guess that's what happens in limited markets were there just isn't enough competition to create truly stellar product. PLC programming suffers the same fate, but I digress.

So, I'm unclear, is this simply stating that the supercomputer can be reconfigured as needed for custom tasks, or is it actually self-configuring? It would be pretty sweet if it could evaluate the task given to it and then reconfigure the FPGA's for optimum efficiency without further human interaction.

[Ed @ Dec 19th 2005 2:05AM]

In the article they say that they have an alliance of FPGA companies creating better software for this. FPGAs are great in their versatility, but that comes at the cost of speed. I'm guessing that their supercomputer will take a long time to reconfigure, then run reasonably fast once configured.

@Patrick Flynn - FLOPS are Floating Point Operations Per Second. It's a measure of mathematical calculation speed. And tera means trillion.

[ch424 @ Dec 19th 2005 2:05AM]

#6, yes and no, really: their objective is to make the supercomputer self-configuring, which would be really, really clever stuff, and as you and #5 say, that would save loads of time _once_ they've got it working!

#4 Are you a mac user by any chance?

[sigh @ Dec 19th 2005 2:05AM]

Supercomputers have 'hundreds of teraflops'?

http://www.top500.org/lists/2004/11/

Big Blue is #1 with 77 TFlops. I know this is a blog, but c'mon now. Let's be a little more professional.

And based off the slides, I think the Cell processor is rated at 220 GFlops (X360 at 110), which sounds a bit more feasible.

[flyboy @ Dec 19th 2005 2:05AM]

The Cell is 220GF, the PS3 GPU is 1.8TF, putting the total system at 2TF. Some people have met success getting GPUs to run non-graphics code, so it may be possible to actually tap that 2TF for useful work (although since there won't be home-brew development, it's unlikely to ever happen).

[Thomas Ricker @ Dec 19th 2005 2:05AM]

#10, your data is 6 months old. Just one example of 100+ teraflops...

http://www.llnl.gov/pao/news/news_releases/2005/NR-05-03-09.html

[elmer @ Dec 19th 2005 2:05AM]

Yeah, bluegene's the only one that's exceeded bluegene's performance so far and they plan to double it again int the next 6 months, but these these are only officiated by the top 500 every year or so. Guess who produced both bluegene and the Cell; that's right, your big blue friends and mine, IBM. Can't wait to hook up a thousand PS3's on gigabit ethernet for a 500,000 dollar 200 teraflop machine (forgetting the RSX) rather than the 100 million they shelled out for bluegene so far - and the bluegene project was cheap. I hear the cells are very flexible.

[Yoda @ Dec 19th 2005 2:05AM]

Overloads Engadget?

You seem to have an obsession with being dominated and controlled by someone or something.

I don't want to make any assumptions but its seems a bit odd...

[isaac @ Dec 19th 2005 2:05AM]

love the terminator reference, that was the first time i ever noticed product placement in movies was when i saw that scene.

[Dg @ Dec 19th 2005 2:05AM]

this is nothing new - the Splash 1 was an FPGA based supercomputer system built in 1988 - and the Splash 2 was running by 1993. Small systems by today's standards - but fairly beefy in their time - http://www.hpcc.gov/pubs/bluebooks/1997/nsa/splash.html

[KK @ Dec 19th 2005 2:05AM]

Supercomputers are rated using *Double Precision* GFLOPS/TFLOPS. CELL as seen in PS3 is only capable of 26 GFLOPS *Double Precision*. The infamous 200+ GFLOPS from CELL is *Single Precision* which is useless for supercomputing.

[Stewert Dent @ Dec 19th 2005 2:05AM]

Thomas Ricker, Nice post, but would anyone really think of the large metropolitan area of Edinburgh Indiana :), unless you are an Indiana Hoosier (Citizen or Student)?