Kutaragi Ken: "PS3's GPU isn't an upgraded PC GPU"

[marconelly!]

Well, that block diagram would indicate that RSC is quite a bit different beast than NV40, or even what's rumored of G70.

For example, what's the shader processor? Is it some kind of scheduler simillar to what it's used in R500? Why are pixel and vertex shaders not specified on the diagram, but just represented as "Vector ALUs" Is that where you guys got the idea that vertex shaders will be doing one scalar + one vec4 insted of 4x scalar like NV40 vertex shaders are doing and what MS tech people assumed? I'd also like to know what exactly are SFU units doing, some documentation on them.

 

[Xmas]

Thanks. I'm curious, do you have any specific links/ documentations on these SFUs?

There's hardly any documentation available about the hardware implementation. But I've read somewhere that the SFU of NV40 is actually split across both ALUs.

EX2, LG2, RCP, SIN, COS, and SCS are the supported single-cycle special functions, POW and RSQ are actually macros that take two cycles:
http://graphics.stanford.edu/projects/gpubench/results/6800Ultra-7189/ (don't mind the wrong results for add, sub and rcp, they're optimized away apparently)

RCP takes place before SU0 AFAIK, the other ops replace the alpha channel of SU1 (according to GPU Gems 2, chapter 30).

I guess the additional NRM_PP could be counted as SFU, too.

Well, that block diagram would indicate that RSC is quite a bit different beast than NV40, or even what's rumored of G70.

For example, what's the shader processor? Is it some kind of scheduler simillar to what it's used in R500? Why are pixel and vertex shaders not specified on the diagram, but just represented as "Vector ALUs" Is that where you guys got the idea that vertex shaders will be doing one scalar + one vec4 insted of 4x scalar like NV40 vertex shaders are doing and what MS tech people assumed? I'd also like to know what exactly are SFU units doing, some documentation on them.

NV40 vertex shader does one vec4 and one scalar/SFU, not 4 scalars. The "shader instruction processor" is most likely responsible for flow control. In NV40 pipeline diagrams there sometimes is a "branch processor".

 

[marconelly!]

NV40 vertex shader does one vec4 and one scalar/SFU, not 4 scalars. The "shader instruction processor" is most likely responsible for flow control. In NV40 pipeline diagrams there sometimes is a "branch processor".

OK, I didn't know that. Why did the MS people used 4 scalars in their calculation then, if they based the prediction on NV40?

*edit* also, what would be the optimal assets distribution in a RSX memory config. Where the textures and vertex objects should be stored, and where the framebuffers, considering it's bandwidth is basically split between the GDDR3 and XDR? Also, isn't it's total bandwidth actually pretty high compared to leaked G70 specs (total of 57GB/s on RSX vs. 34GB/s on G70)[/quote]

 

[j^aws]

@Xmas, thanks for links and info...

NV40 vertex shader does one vec4 and one scalar/SFU, not 4 scalars. The "shader instruction processor" is most likely responsible for flow control. In NV40 pipeline diagrams there sometimes is a "branch processor".

OK, I didn't know that. Why did the MS people used 4 scalars in their calculation then, if they based the prediction on NV40?

Who are these MS people? Please don't tell me it's Major Nelson and co... :?

...*edit* also, what would be the optimal assets distribution in a RSX memory config. Where the textures and vertex objects should be stored, and where the framebuffers, considering it's bandwidth is basically split between the GDDR3 and XDR?...

Dave recently posted,

Talking to one developer recently about the memory split on PS3 he said that they would generally budget around a 50%/50% split between system and graphics memory anyway, so its probably the exception rather than the norm that developers would place texture data outside of graphics RAM anyway. There is 256MB on the graphics for a reason, its not there not to be used.

http://www.beyond3d.com/forum/viewtopic.php?p=523971#523971

So the 50/50 % split sounds reasonable...

...Also, isn't it's total bandwidth actually pretty high compared to leaked G70 specs (total of 57GB/s on RSX vs. 34GB/s on G70)

Where did you get the G70 numbers from?

 

[Nite_Hawk]

Dave recently posted,

Talking to one developer recently about the memory split on PS3 he said that they would generally budget around a 50%/50% split between system and graphics memory anyway, so its probably the exception rather than the norm that developers would place texture data outside of graphics RAM anyway. There is 256MB on the graphics for a reason, its not there not to be used.

http://www.beyond3d.com/forum/viewtopic.php?p=523971#523971

So the 50/50 % split sounds reasonable...

Dave made a statement like this in one of the other threads, though Fafalada disagrees:

he mentioned that they would generally budget RAM allocation at 50% for system and 50% for graphics previously; not sure that is representative of all developers.

It's most definately not representative (at least on consoles). Besides if you want to go just by "one developer" examples, I've once heard of an app that used 15MB of main memory for just strings...
...
on a PS2... :?

http://www.beyond3d.com/forum/viewtopic.php?p=529580#529580

Nite_Hawk

 

[marconelly!]

Who are these MS people? Please don't tell me it's Major Nelson and co... :?

Yeah, those are the only MS people who did the analysis that I know of I already said I know the whole thing is a massive load of BS, but that one aprticular calculation seemed believable. Now even that seems very shaky in the light of the info posted here.

Where did you get the G70 numbers from?

They are all over the place. Take a look here:
http://www.nvnews.net/vbulletin/attachment.php?attachmentid=11640
(it actually says: Memory Bandwidth 38.4GB/s, so I didn't memorized the number correctly)