New supercomputer is a rack of PlayStations | Hacker News Reader

New supercomputer is a rack of PlayStations

smh.com.au

16 pointsauferstehung12 comments

The PlayStation3 has gigabit Ethernet, which is ideal for multi-tier servers. However, the PlayStation3 only has 512MB RAM. Regardless, it may be useful for many computationally intensive tasks:

The PS3's hardware has also been used to build supercomputers for high-performance computing. Terra Soft Solutions has a version of Yellow Dog Linux for the PlayStation 3, and sells PS3s with Linux pre-installed, in single units, and 6 and 32 node clusters. In addition, RapidMind is pushing their stream programming package for the PS3. Also, on January 3, 2007, Dr. Frank Mueller, Associate Professor of Computer Science at NCSU, clustered 8 PS3s. Mueller commented that the 512 MB of system RAM is a limitation for this particular application, and is considering attempting to retrofit more RAM. -- http://en.wikipedia.org/wiki/PlayStation_3

Holy...when I was working at my last job (a financial software startup), I joked to my boss that we should port our software to the PS3's GPU and run a cluster of PlayStations. The math in quantitative finance is rather similar to the sort of massively-parallelizable vector calculations in computer graphics. Plus, I had this amusing image of hedge fund managers slaving away at their Dell workstations while the PlayStations were all locked in the server room, crunching numbers.

I'm impressed that someone actually did it, which goes to show that there's no idea too crazy to implement.

Hampster powered workstations... I'm not sure if it would be a disaster for productivity because they're cute or because their poop stinks, and to power a workstation the problem would be multiplied to the 2nd or 3rd power.

I'm sure someone could implement it, but I'm not sure how wise it would be... maybe a PetSmart could do it first.

Calling 16 PS3s a supercomputer is a bit of a stretch. The article doesnt quote any numbers either. Cell with 8 SPEs can do about 200 Gflops SP matrix multiply. So I assume PS3 can do about 150 Gflops. 16*150=2400 gflops single precision .. not a supercomputer by todays standards i would say.

Certainly 16 PS3s is no Blue Gene. But perhaps good enough for a lot of problems and hopefully cheaper to build than using standard desktop/server hardware (since you get the PS3 hardware at a discount).

It's still a lot faster than your average physics department's 32-node Xeon cluster.

I'm curious - where did you get that 200 Gflops number from? Is that theoretical, or actual for large values (say 10k) of N?

Thats from a paper called "Potential of Cell processor for scientific computing" .. they quote a figure of 200 gigaflops for single precision GEMM (generalized matrix multiply) for large square matrices (using a simulator?). Assuming their model is very accurate, its actual 200 gflops.

Wouldn't it be easier to use "full" Cell processors? Are they much more expensive than PS3s? I assume IBM sells other computers with Cell chip?

How do cell chips compare with modern graphic cards, I think they can be abused to do a similar thing?

According to http://www-132.ibm.com/webapp/wcs/stores/servlet/CategoryDis... a blade with two cell chips costs $10k. Two PS3s come in quite a bit cheaper than that. Okay, so they don't have as fast an interconnect and as much RAM, but it depends on your problem whether that matters at all.

Modern graphics cards are also useful for number-crunching, but it's pretty awkward to develop general-purpose code for them. Not that the Cell is all that easy to code for with its SPE-local SRAM and explicit DMAs for memory access. As far as I can tell, though, the Cell development tools are more mature, and it wins in terms of FLOPS-per-watt. I suspect the real win for the Cell is that its interconnect latency is tiny compared to your average GPU, which is great for clustering.

The thing that's killing mainstream adoption for both options in the HPC sector is lack of double-precision (64-bit) floating-point support. There's a version of Cell on its way that fixes that problem, however.

Note: I've not written code for the Cell myself, although I have for GPUs.

Thanks, very interesting.

sounds like a business venture

Reply on news.ycombinator.com