[leechan25]

Why wouldn't Sony use only cell based chips for PS3 primary func tions? It looks to me like rather than reinventing the wheel, their simply getting the best player in their respected industries to apply their knowledge to customizing cell base chips for PS3. I mean you look the partnership with Nvidia, their not designing their own GPU - their use cell as a blueprint for PS3 gpu. Why not for sound related chips and more?

We know that cell is designed to increase in processing power when connected with other cell based products. Why would Sony put other chips into their hardware at won't be as compatible? Plus future fab cost on non cell based chips can't really be controlled? Sony would then be pulling a Microsoft when it comes to future production. It's not needed. Plus Sony's trying to build demand for cell based chips. Why promote someone else?

[Guden Oden]

Most probable reason would be that general processors, even multi-core designs like Cell, simply isn't particulary suited for rasterizing 3D graphics. You'd either end up spending an equal amount of transistors as a dedicated 3D chip and get (often hugely) inferior performance, or have to spend a vastly higher amount for a similar performance level.

You use the tool that is best for the task, quite simply.

[akira888]

One computational task that's really cheap to hardwire but expensive to calculate on a general purpose chip is texture filtering. Standard bilinear alone takes (assuming we have complemented inputs for free) 4 scalar multiplies (or one scalar multiply with 4 scalar subtracts), 1 4-vector multiply, and 3 4-vector multiply and adds. This is why almost all software real-time rendering engines lack everything more advanced than dithered point sampling.

Effective management of the necessary context switches due to the 40-80 cycle latency of fetching texture data is also be difficult to imagine. Needless to say moving fragment shading onto general purpose CPUs seems rather distant at this point, if ever.

[akira888]

edit: argh, double post, sorry guys.

[version]

raytracing computing on cpu without gpu is possible,
but must have fast paralell random access and large memory

[Megadrive1988]

Originally Posted by leechan25

Why wouldn't Sony use only cell based chips for PS3 primary func tions? It looks to me like rather than reinventing the wheel, their simply getting the best player in their respected industries to apply their knowledge to customizing cell base chips for PS3. I mean you look the partnership with Nvidia, their not designing their own GPU - their use cell as a blueprint for PS3 gpu. Why not for sound related chips and more?

We know that cell is designed to increase in processing power when connected with other cell based products. Why would Sony put other chips into their hardware at won't be as compatible? Plus future fab cost on non cell based chips can't really be controlled? Sony would then be pulling a Microsoft when it comes to future production. It's not needed. Plus Sony's trying to build demand for cell based chips. Why promote someone else?

it makes no sense to do rasterizing on Cell's PUs+A|SPUs. that's why Nvidia was selected because they have an enormous amount of engineering resources and IP in this area. Nvidia is basicly the other part of SGI that did not go to ATI, plus various acquisitions from other 3D graphics companies outside of SGI. basicly, outside of E&S, PowerVR, XGI and a few other small companies, the 3D graphics world is divided into 2 camps: ATI and Nvidia. both are more or less equal. Sony went with one camp. Nvidia is going to provide the rasterizing end (at least) of PS3. whereas Cell would not be suited for these tasks.

[version]

[Wunderchu]

Originally Posted by version

that's quite cool ... thanx for the graph, version 8)

[london-boy]

Where's that graph from? Doesn't explain much.

[nAo]

the graph just shows the the algorithmic complexity of various rendering techiniques. Linear for 'standard' scan conversion and logaritmic for ray tracing.
Deferred rendering is shown to be more efficient than IMR, in fact it has a smaller than IMR angular coefficient.
Notice also that it's not shown the constant cost of each approach.
IMR should have the lowest constant cost, ray tracing should have the highest one. Costant cost of each implementation is important cause it helps us to understand when an algorithm become 'better' than another one.
Ray Tracing with its logoratmic complexity should win when we are dealing with a high number of primitives..the graph clearly shows this tendency.

ciao,
Marco

[london-boy]

Yeah Marco, i got that, silly!

It just doesn't tell me much of the point where ray tracing starts being relatively cheaper than the other 2. In reality. I guess no one knows.

[j^aws]

Originally Posted by version

For the two intersections with deferred and poly-stream with RT, what computation power and Polys are you estimating for those :P