Slashdot Mirror
Eight PS3 'Supercomputer' Ponders Gravity Waves
Jamie found a story about a inexpensive supercomputer being used by an astrophysicist to research gravity waves. The interesting bit is that the system is built using 8 PS3s. Since nobody is actually playing games on the system, it makes sense to use them for research projects like this, but I really wonder now what is defining 'Supercomputer'... I mean, a hundred PS3s sure, but 8? I think we are de-valuing the meaning of the word 'super' :)
It's only six cores available to Linux per Cell processor on a PS3. One is reserved for the Game OS, and one is disabled to achieve a higher yield on fabrication. (The Game OS is always running, since Linux actually runs on top of a hypervisor.)
8 PS3s gives you 1.2 teraflops of single-precision performance or a similar number if you stick to integer operations (6 SPUs/PS3 gives ~150 GFlops). 1.2 teraflops is a supercomputer in my book.
:)
Using Jack Dongerra's single-precision algorithms that do half the work in single and the other half in double precision, you can maintain a high level of performance and precision. And, the unique architecture of the Cell opens up some interesting algorithmic research issues, allowing scientists to publish twice for the same work: once for the science results, once for the computer science results.
On the flip side, the Gigabit ethernet on the PS3s isn't really 1GB - the PPU can barely keep up. So, extra care must be taken around communication points. And, a similar Intel/AMD-based rack would run about $20k and is much easier to develop for, so if your labor is expensive (i.e., you're not in academia), PS3 clusters may not make much sense.
-Chris
Sony's estimate of 1 teraflop per unit is more than a little bit optimistic, and also counts all the operations performed by the graphics system, which the average researcher won't have access to. For a slightly more realistic estimate of what 8 ps3's running linux can do, I'd point you towards http://www.terrasoftsolutions.com/store/index.php?submit=software&submitimg%5Bhardware%5D%5Bsolutions%5D=1, who sell turnkey ps3 clusters. They are claiming that 8 units together break the theoretical teraflop mark, which seems more realistic to me. While that isn't among the fastest computers in the world by any stretch, it's still solidly in the realm of what I'd call a supercomputer.
The cell is a fantastic piece of equipment - Dr. Dobb's has what I think is an excellent analysis of the kinds of performance benefits that it offers at http://www.ddj.com/hpc-high-performance-computing/197801624. I'm currently running one at home in a PS3 (for neural networks that drive an AIBO - I love Sony's tendency to dump hugely expensive hardware at mass-market prices), and I have every intention of picking up more used ones over the coming months to cluster together as the networks continue to grow. Even all by it's lonesome with code that's far from optimized, the one I have is running about 10 times faster than my main desktop for roughly equivalent computations.
(Note that your mileage may vary - I just happen to like playing with systems that parallelize really well)
How much something "costs" to produce isn't as simple as the cost of parts. It may cost $5 for just the parts, it may also cost and extra $1-5 for the direct labor to put the board together, and it may cost and extra $1 to ship it. Then there's the cost of maintaining a chip fab, indirect labor and mangement costs and bribes to the chinese government to keep their cheap wage factory certified, divided over the 2-300 PS3's produced per month to keep up with worldwide demand, not to mention the cost of worldwide marketing, the cost of years of system & Blu-Ray R&D spread over each unit... keep in mind that the cost of making your PS3, Wii or software title is more than just more than the sum of the cost of the individual parts.
Sorry, that was my Cost Accounding class talking, I'll stop now.
It would've been cheaper to just buy a Cray.
If you read the article and followed the link to his PS3 Gravity Grid site, you'd know a couple things about the cost (FREE) for this computational power:
#1) The total cost of purchasing an entire "PS3 Gravity Grid" supercomputer for yourself is less than the cost of a single simulation run on a BlueGene. In other words, you can buy the cow, the pasture, and a barn for the price of a gallon of milk.
#2) Sony *DONATED* his 8-node cluster (albeit with 20GB PS3's which they were closing out at the time) so he actually got a "supercomputer" for nearly free.
#3) The power of the 8-node PS3 cluster is roughly the same as a 200 node partition on a BlueGene SuperComputer (1 PS3 = 25 Blue Gene nodes). With 8 Cell CPUs, he has 56 SPU's running at ~3GHz to crank through his computations. This would mean a single CELL SPU is roughly 4X more powerful than a single BlueGene node which isn't unreasonable considering that it runs at a higher clockspeed (the supercomputer has to worry more about heat dissipation with hundreds or thousands of cores).
#4) I believe that by the US Gov't's somewhat outdated standards, a PS2 qualifies as a supercomputer. The FPU power in a PS3 is on ther order of 200 times higher than that of the PS2 for single precision and considerably more for double precision (which is emulated in software on the PS2).