Shader Model 4.0 -- what do we know about it?

R500 (which will be found in XBox2) is already WGF 2.0 compliant. It will have 48 ALUs (unified shader units) in the chip. These ALUs can perform Pixel Shader, Vertex Shader and Geometry Shader calculations, the assignment will be load-balanced by the driver. AFAIK nVidia won't follow this direction and doesn't converge the two shader units physically. Of course they are forced to use the unified shader model on driver level, but they can dedicate a type of ALU for pixel, a different type for vertex, and a third type for geometry processing. Only time will tell which solution is better.

Unificiation on hardware level (ATI's approach):
+: Flexibility, better load-balancing
-: some performance penalty due to the lack of optimization

Unification only on software level (nVidia's approach):
+: Performance gain due to the optimization of the hardware units
-: Can be bottlenecked by any shader unit if the stress on the different ALUs are not in balance

 

TBH, I think SM4 is going to be another step down the line of GPU's becoming little RISC processors with more and more flexibility.

ATi are already talking about separating the texture fetch/etc. sections of the pipeline and living with an abritary number of ALU's running in parallel for PS or VS calcs. This looks like the way its all going to end up IMO.AS for Nvidia suggesting they wont go the multipurpose vs/ps unit way - well, how far has arrogance got any other company in 3D graphics? COUGH 3Dfx COUGH

 

Why do you think the path we believe nvidia has taken is driven by arrogance?

 

I don't think RISC would be an accurate description (particularly since it describes the instruction set, which is secret).

This doesn't necessarily have anything to do with arrogance. nVidia may simply feel that there are some significant problems with a unified architecture, enough such that they can make better products without one.

 

Or incapable of making a unified approach at present .

 

I think Nvidia will go more in the direction of having equal functionality in PS/GS/VS units, however each unit will be optimised differently. So while a VS can be forced to function as a PS, there will be a significant performance penality. That seems to make the most sense imho.

 

I don't think there is any arrogance involved here.....

Nvidia's argument on the unified shader thing is just as valid as ATI's IMO.

All they are saying is that while both units have similar functionailty the mix of instructions in the short term will be radically different (more texture reads in the PS vs more ALU ops in the VS).

If this is true (and I'm not saying I even agree with it) then it's possible that by optimising the shader units for their jobs you could end up faster than a unified design with the same transistor budget.

 

ATI are presently using the same argument - Xenon graphics looks to be unified at the hardware level whilst R520 looks to have discrete units; why? Because the workload for the PC currently are different from those in a closed box environment with titles built specifically for the hardware at hand. The question is whether ATI are too early when they bring it to the PC or whether NVIDIA will feel the need to go that route in the future.

 

Wouldn't it be more fair to say the R500 just isn't ready yet and the Xbox is a nice testing ground for a unified architecture considering you will be working with MS for the duration?

I'm just saying it might not so cut and dry, but I don't pretend to know anything so it's just a question.

 

Maybe Dave's just saying that Xbox 360 isn't going to be CPU-limited like PC and the devs will use that to their advantage...

Actually, isn't one of the major goals of WGF to massively reduce the DX overheads?

Jawed

 

PC's are only as CPU-limited as the software demands. Things only seem to be more CPU-limited now because software developers expected a more rapid pace of CPU development.

Similarly, since the software will be made directly for the next X-Box platform, well-written games will be every bit as much CPU-limited as they are GPU-limited on that platform.

 

They certainly should be at least.

 

Hmm..you sure? I was under the impression it was taking some elements from WGF 2.0 but not necessarily all. I thought WGF 2.0 was going to bring SM4.0 into the mix, which the R500 won't be compliant with? Or have things shifted..WGF 2.0 scaled back to match R500 capabilities? WGF 3.0 to be the "real" next DX?

Is WGF2.0 even nailed down yet?