nvidia:rsx complete.

Well, Xenos' daughter die suposedly spends 20-25 million on its 8 ROPs. They are of course both simpler (fixed AA sample pattern, limited to 4x, no FP16 blending, etc) as well as more capable (relative to most ATI parts excluding RV530) (double z and 4x samples per clock). So 16 ROPs, depending on how expensive the 4 sample ability or FP16 blending are, could amount to 40+million transistors. (I think. Somebody check everything I said. Stat!)

As far as the fragment pipes, though. Only 12-16 seems way too extreme. 16 is hard to buy. If there's redundancy, I wouldn't expect anything less than 5 complete quads in there.
 
What do you guys think Sony and Nvidia will do with the extra space the RSX will have, being that some of the stuff the GTX7800 has (I forgot the technical terms) for like TV of something? I read that the RSX doesn't need alot of the stuff that a normal PC GPU has.

Want can Sony and Nvidia replace that extra space with?
 
since RSX has a 128-bit bus, I'm guessing only 8 to 12 ROPs and 12 to 16 fragment pixel shaders will be active ?
that might account for the redundancy.

Umm..this would leave the PS3 very weak so no.

Not so much the ROPs as X360 has only 8 and they are not the bottleneck, but 12-16 pixel shaders is in league with a current mid-range card.

I'm not sure what they'll do but expect 20 pixel pipes minimum, probably 24, possibly >24. That is basically trhe most important area of how good the games will look onscreen.
 
TurnDragoZeroV2G said:
Well, Xenos' daughter die suposedly spends 20-25 million on its 8 ROPs. They are of course both simpler (fixed AA sample pattern, limited to 4x, no FP16 blending, etc) as well as more capable (relative to most ATI parts excluding RV530) (double z and 4x samples per clock). So 16 ROPs, depending on how expensive the 4 sample ability or FP16 blending are, could amount to 40+million transistors. (I think. Somebody check everything I said. Stat!)

As far as the fragment pipes, though. Only 12-16 seems way too extreme. 16 is hard to buy. If there's redundancy, I wouldn't expect anything less than 5 complete quads in there.
Xenos has 8 color ROPs, but 64 Z ROPs (including 4x AA).
 
Sorry to be slightly off-topic, but am i the first to congratulate and welcome back our beloved Acert93?

Welcome back!!


EDIT: Just checked, and no i'm not the first one.... oh well, welcome back!!
 
TurnDragoZeroV2G said:
So 16 ROPs, depending on how expensive the 4 sample ability or FP16 blending are, could amount to 40+million transistors. (I think. Somebody check everything I said. Stat!)
Without even disputing your numbers, I'll just add that increasing the ROP count would require a complete reworking of the eDRAM array itself, the bandwidth it provides is balanced against the performance of the rasterizing section of the die, and increase the hunger for data they'd also need to increase the chip's ability to satisfy it. This would obviously increase chip complexity and cost...

Besides, pure pixel fill isn't going to be a bottleneck on the 360. 4 billion pixels (peak) is enough to fill several dozen 720P screens per frame at 60 frames/sec. More ROPs won't be a neccessity.
 
Guden Oden said:
Without even disputing your numbers, I'll just add that increasing the ROP count would require a complete reworking of the eDRAM array itself, the bandwidth it provides is balanced against the performance of the rasterizing section of the die, and increase the hunger for data they'd also need to increase the chip's ability to satisfy it. This would obviously increase chip complexity and cost...

Besides, pure pixel fill isn't going to be a bottleneck on the 360. 4 billion pixels (peak) is enough to fill several dozen 720P screens per frame at 60 frames/sec. More ROPs won't be a neccessity.

I think he was answering my question as to how many transistors ROPs typically take up in regard to G70/RSX (by generalizing from Xenos' ROPs) .

Since it's been in question as to whether RSX would even need 16 ROPs, it might be something that could have been reduced... more speculation!

If we take out 4-8 ROPs and some stuff like PureVideo there is quite a bit of room for saved transistors (which was the heart of the question -- if transistors for that stuff was taken out, and the transistor count was still ~300m... what replaced the stuff?).
 
Megadrive1988 said:
since RSX has a 128-bit bus, I'm guessing only 8 to 12 ROPs and 12 to 16 fragment pixel shaders will be active ?
that might account for the redundancy.
There's already redundancy for Shaders Quads on the G70.
 
Guden Oden said:
Besides, pure pixel fill isn't going to be a bottleneck on the 360. 4 billion pixels (peak) is enough to fill several dozen 720P screens per frame at 60 frames/sec. More ROPs won't be a neccessity.

Actually fillrate is the biggest C1 problem, as can be seen on any games so far - none is using AF and only bilinear filters!
 
Actually fillrate is the biggest C1 problem, as can be seen on any games so far - none is using AF and only bilinear filters!

Might just be me but Anisotropic Filtering, Trilinear Filtering etc. use Texel fillrate and not Pixel fillrate.
Some Xbox 360 games are just not using high AF because they were totally rushed.
 
Nemo80 said:
Actually fillrate is the biggest C1 problem, as can be seen on any games so far - none is using AF and only bilinear filters!

I guess the problem with an architecture overly focused on pixel shaders. Looks like G70 based PS3 will fare better in this regard.

Xbox 360: 16 TMU's x 500 MHz = 8,000 MPixels/sec.
PS3: 24 TMU's x 550 MHz = 13,200 MPixels/sec.

A difference of 1.6 times just on # of units, and clock rate. The G70 has 24 KB of texture cache per pipeline, and wonder how much locality Xenos has concerning that?
 
Anisotropic filtering consumes texture bandwith, not fillrate. Both RSX and Xenos seem to have about the same amount of that, as the PS3's GDDR bus will be mostly consumed by framebuffer access - which leaves the XDR bandwith, and that's almost the same as the main RAM bandwith in the Xbox.

Of course the higher complexity of the PS3's memory systems also leaves more room for developers to experiment. But as of now, what I've described seems to be the general approach...
 
Nemo80 said:
Actually fillrate is the biggest C1 problem, as can be seen on any games so far - none is using AF and only bilinear filters!
1) Using AF over bilinear texture filtering would actually lower fillrate demands because the pixel shaders would take longer and Xenos has dedicated fill bandwidth.

2) The texture cache on Xenos is smaller than on other GPUs, at least that's what that one quirky Japanese engine/game demo concluded. If true, developers would stick with trilinear filtering over AF because they're used to putting everything into one shader instead of multipassing.

3) It is possible for Xenos to be fillrate bound, but it's not too likely AFAIK. If you dedicate every shader to pixel shading, your pixel shaders have to take less than 6 cycles to bottleneck at fillrate. Not many pixel shaders are that small anymore and most draw commands require some shaders for use on vertices. Also, I'm not sure you'd be better off with a more conventional GPU because the ROPs would probably not be 100% utilized for the small shaders we're talking about. This is especially the case with more involved rendering like alpha blending.
 
Edge said:
TMU: Texture Mapping (or management) Unit
Oh, thanx.

Well, I think 360gpu has 3 simd units with 16 shaders per unit, making total 48 shader processors, with each shader made from 4 alu; each can be either vertex or pixel shader. So where is sweet 16 (TMU)?

are you talking about fetch units?
 
Last edited by a moderator:
Back
Top