(Untitled)

[jzig]

I got into an argument over IRC about programming the PS3 vs. programming the 360. Most of my arguments are based on second-hand information or old articles. Anyone read anything recent that discusses the kind of operations that are efficient to run on the cell's SPE? I read somewhere that most PS3 games only use one or 2 spe's, but I can't source

Comments

I have never even seen a PS3 outside of its box, but

[kvance]

All I know is that the SPEs have no branch prediction, making them useless for AI. I think they would be pretty good for graphics post-processing, which is code you'd have to rewrite on each platform anyway. Probably be good for physics too, if you had your own physics engine and the resources to port it.

Re: I have never even seen a PS3 outside of its box, but

[wm_eddie]

The SPEs do prediction (in the sense that they always put a branch in the pipeline and don't just stall until it's time to test). They always predict false, unless you tell them at least 7 instructions before hand to predict true. So they probably suck for AI but they could be good for filling out bayes tables.

[coderanger]

I never said anything about cross-platform, that would just suck. To take decent advantage of the Cell's hardware you would have to go PS3-exclusive. I'll have to do some more research before I can back up my random claims of usefulness though :-)

[wm_eddie]

With GCC, the SPEs can access the PPE's address space (and vice-versa) by using the type qualifier __ea. This does some software caching too so by using this method you can create two functions that on the surface work in the same way as having multiple cores. But unless you can use the caching effectively (which you can vary at compile time starting from 8KB) the overhead will outweigh the any benefits. Since I don't have a PS3 I can't benchmark it to see how big the overhead is though.