More info about RSX from NVIDIA

[nAo]

First G70 review is out, I found this nice recap:

http://www.hardspell.com/hard/showcont.asp?news_id=14372&pageid=3105

I speculated pixel shader floating ops per clock cycle on RSX would be the same number of floating point ops Xenos ALUs can do, it seems this is true.
(I'm speaking about programmable flops)

Google translation from chinese to english:

Sony's PS3 main engine used based on G70 the overhead construction design RSX graph core, ncVidia official first time produces G70 and the RSX detail requirements contrast, simultaneously also provided the R500 graph core reference data which Xbox used.

 

[j^aws]

I can't see the RSX pixel column in my browser! :?

 

[pc999]

~400Gflops/s :? , is this real

How many of them programable

 

[London Geezer]

~400Gflops/s :? , is this real

How many of them programable

I guess those are the programmable ones, since the total for RSX is allegedly 1.8Tflop and the total for Xenos is almost 1Tflop.

The 400Gflops/240Gflops might be the programmable part.

I still smell fish though.

 

[passby]

The funny part is, according to the site, that diagram *including* R500 details is provided by Nvidia.

Lots of Chinese terms I don't know how to translate to the English equivalent, so I guess we'll just wait for the english sites' coverage.

 

[nAo]

I was wrong..RSX pixel shading flops per clock is even more than what I was expecting.
They're quoting 27 flops per pixel pipelines, I would like to know how they're counting nrm flops.

 

[jvd]

400Gflops/s , is this real

How many of them programable

more believeable than 1.8 tflops from the confrence .

 

[mckmas8808]

400Gflops/s , is this real

I not the most technical person in the world so I'm asking, shouldn't it have a higher programmable flops per second due to the higher MHz? %550 compared to 500? :?

 

[j^aws]

400Gflops/s , is this real

How many of them programable

more believeable than 1.8 tflops from the confrence .

The 1.8 TFlop was NOT a programmble Flop metric, the same for X360 with 1 TFlop...

 

[London Geezer]

400Gflops/s , is this real

I not the most technical person in the world so I'm asking, shouldn't it have a higher programmable flops per second due to the higher MHz? %550 compared to 500? :?

Yeah but a 100% increase in Flops with a 10% increase in clockspeed???

 

[blakjedi]

To understand performance then, do you sum all pixel and shader performance? For example can the RSX do 5 vertex and 10 Pixel operations per ALU simultaneously giving it 15 ops/ALU whereas the xenos only does five?

Please help me understand the graphics pipeline process.

 

[passby]

Something new -
TSAA Transparent S(?) Anti Aliasing
TMAA Transparent Multi(?) Anti Aliasing.

Demo here
http://www.hardspell.com/hard/showcont.asp?news_id=14372&pageid=3116
1st 2 pics have anti-aliasing off.
2nd 2 pics have 4x AA.
3rd 2 pics uses the new aliasing technique.

Last 2 pics are ATI X800 without AA, then with 4x AA.

 

[Fafalada]

How many of them programable

All of them.

Anyway, I am starting to get worried that we'll really get a bunch of useless video crap transistors still in RSX(given the similarities in transistor counts for G70)...
replace it with sound or something if you can't think of a way to use it dammit. :?

 

[Shifty Geezer]

This is nVidia info, right? So these RSX specs are direct from nVidia? So we can probably discount their R500 (XB360) numbers, not least becuase there is no R500 in XB360!

Certainly here RSX is little more than G70, but what's the tranny count for G70? How much more silicon if any is going into RSX that we have to account for extra functionality (if any)?

 

[jvd]

Something new -
TSAA Transparent S(?) Anti Aliasing
TMAA Transparent Multi(?) Anti Aliasing.

I believe the r500 is capable of this too , I don't know if the r520 is .

 

[passby]

<OT>
I like Faf's new sig.
</OT>

 

[pc999]

Well, if they really want to do 1080p, with the same qualitity per pixer (instructions per shader), they would need almost 2X more power right

 

[_phil_]

TSAA Transparent S(?) Anti Aliasing

sorting (?)

 

[j^aws]

nAo,

RSX ~ 56 vec4 units ~ 56 DOT4/cycle and not 52 DOT4/cycle then!

 

[Rockster]

That chart should show vec3 + scalar or vec4 in the pixel shader, and not vec4 + scalar.