Cell & games

[tongue_of_colicab]

Ive read discussions about this on other forums for some time now and used the search to see if I could find anything about this on b3d but I couldnt really find anything.

What I would like to know is if cell really is that good for games as most people seem to think it is? what are its pro and cons against a more normale cpu like in the x360 or in pc's? What I understand of it cell is great at thing like physics and video encoding/decoding but is it better than a normale cpu? I heard it can is quite hard to make all the spe's work together perfectly and that (alot) of speed is lost because of spe having to talk and share data with each other.

But I dont know wheter this is all correct or not.

 

[Arwin]

I went through part of this myself, on this page, where I learnt a lot:

http://www.beyond3d.com/forum/showthread.php?t=33973&page=6&highlight=cell+rsx

 

[Techno+]

well, the performance of Cell depends more or less on the nature of the applications. The SPEs are deeveloped to take care of single precision floating applications, in a game, single precision calculations include gfx, AI and physics. The SPEs are excellent at AI and physics, but only good at vertex shading, in the case of graphics, so it will be more efficient to use them in generating data structures for the GPU.

 

[tongue_of_colicab]

I went through part of this myself, on this page, where I learnt a lot:

http://www.beyond3d.com/forum/showthread.php?t=33973&page=6&highlight=cell+rsx

Ive read the last page but that seems to be more about cell & gfx but im more interrested in what makes cell better/worse for a console than a more traditional cpu design.

 

[London Geezer]

Isn't this the wrong forum for this?

And hasn't this been discussed to death in the forum this thread should be in anyway?

 

[Rolf N]

<...> its pro and cons against a more normale cpu like in the x360 <...>

Nooooooooooooooooooooo

Cell is decently optimal for software that uses a single/few "control" threads and farms out lots of throughput-oriented work to worker threads. The PPE is competent enough at the first task, and the SPEs are really good at the second kind of task.
The PPE is a relatively simple processor, where tradeoffs have been made to keep the die area down. A "normal" PC processor core (Athlon 64, Core 2) will kill the PPE on spaghetti code.

The SPEs, though they are complete processors, have a limited view of memory and strongly emphasize SIMD performance. The Local Store is a way to get rid of the tremendous synchronization complexities you'd have when you couple so many cores together and try to make things coherent. It makes external memory access more difficult, but at the same time it makes access to the core's working set much faster than would be feasible with a "normal" cache. Local Store latency and bandwidth are very good, which is useful for keeping the ALUs busy.

A million write-ups exist that go into more detail. Also I'm not in the slightest interested in telling you what's better than what.