Cheapest way to program a cell processor

[_D_]

What the hell does IBM know anyway?

Windows=

[Ducc]

Hm. What would that be good for? Maybe some kind of graphics/simulation server? Could be cool.

[_D_]

I can only tell you how I'd use it. We'd create the highest fidelity environmental model possible to model our toys before testing them. Instead of relying on a monte carlo simulation which uses random number generators to simulate subtle changes in the environment, I'd use a stochastic approach that actually generated discrete numbers from a measurement.

We've reached the limit of usefulness for traditional discrete event simulations. The only way to make our models more useful is to stop making isometric assumptions. For example, the sound/velocity profiles (SVP: how fast sound travels at a given depth of water changes drastically because the surface will be a different temperature than deeper water) of the ocean are assumed to be uniform over a given area. That works for a sufficiently small area. When an operational area becomes large, then the SVP varies. The better you can account for these changes, the better you can model the behaviors of your sensors.

[Ducc]

Right, but is that going to have a much better price/performance than a dual core Intel chip?

[_D_]

Can't compare the two, they are too different. Each cell processor has 7 vector units. The best analog on a Intel chip is the SIMD registers. For reasons unknown, Intel has refused to add real vector processing to the Pentium line. A better comparison would be to a MIPS processor from SGI.

In addition to CFD, vector units are perfect for decoding large media files. Any array processing can be done with near N operations with a vector unit. Incidentally, this is why Apple wanted the Altivec.

It's not that we want to get better performance than a Intel CPU and save a few hundred dollars. If you include the development time, that wouldn't make sense if it would take two developers 3 months to port our sim. The ROI would be too great.

We want to approach SGI performance and the cost savings per node would be orders of magnitude less. Even if we spent $250k for a 3 person development team including hardware, it would still be less than one node in our Origin farm.