Mintmaster said:
I'm sorry I came off that way. I was responding to a post about Cell taking over vertex shading duties. Okay? I have no doubt that there are a million other non-vertex-shading things that Cell will excel at, and a million new techniques that are enabled by Cell (and Xenon too, for that matter, which is no Cell but is still big step up from a Celeron).
And your response, IMHO, was innacurate.
You mentioned bandwidth and claimed it to be a restriction - but firstly your numbers were a bit extreme (you're talking about the numbers of "iterators" in most cases, which is fair enough - even if you don't always need as many as you picked as an example - but your actual *bandwidth* calculation seems to be assuming 32-bit float vector-4's for everything... which is definitely excessive), and secondly even at the kind of complexity you were talking about, it doesn't prohibit Cell from helping. Personally I'd be willing to bet that given a scene with 2M polys, and the same scene with only 1M polys but 2-3 times the processing power available to calculate their attributes, the lower-poly scene will look better - unless that power is squandered.
It's the same reason a lot of devs are arguing the case for continuing to use lower-resolutions rather than have to render in HD - they'll be able to spend longer on pixel-shading. Cell isn't likely (again, IMHO) to aid much in the way of pixel-shading, but it can definitely help out on the vertex side.
As I've also said, if you have a shader (pixel or vertex) that requires a large number of inputs, then indeed I'd expect your performance to take a hit - even with infinite bandwidth, attributes cost cycles, and the shader will have to do something with the inputs anyway, which is also likely to chew cycles. So you mentioning bandwidth was a spurious point because you're going to be bitten everywhere else as well - probably more seriously than the bandwidth. So why did you bring up the bandwidth?
No-one I've seen here has discussed Cell *replacing* the RSX vertex shaders - why would it? In X360, offloading vertex work would allow the GPU to spend it's unified resources on pixels. But RSX has dedicated shaders - why leave them idle?
And now you're brining up RSX *controlling* the SPEs?? Where did that one come from?
You seem to be arguing against an imaginary viewpoint... it's easy to pick an extreme and polarised view and then prove it wrong - but in reality, other than a few random suggestions that get chewed over here every now and again, no-one is seriously suggesting the kinds of things you seem to be taking issue with.
No, the points being made are about Cell *assisting* the vertex shading. In my mind that means Cell does some of the work, and the VS units on the GPU does some more - as a two stage process, not as some kind of multi-tasking effort. I'm doing stuff along those lines, so naturally I'm going to say that it's a reasonable suggestion.
Look, I'm not interested in wasting time arguing the minutae of your posts. Here's the point I'm really making:
Cell can and will be used to
assist in vertex processing in the PS3, because doing so allows more complex vertex work to be accomplished than RSX could do on its own. Bandwidth will not typically be a limiting factor in this.
If you disagree with that, fair enough - it's tough for me to offer any concrete proof that I'm already doing this kind of thing, so your word is just as good as mine. But so far I've seen you offer nothing that actually prevents this from happening (I wonder if my code will magically stop working if you do?)...
