Official speculate before its too late: RSX Spec thread

[Shifty Geezer]

Actually, though I appreciate Version has a particular role to play in this forum and he does it well, aren;t the rules and regulations such that he ought to add meaningful comments and insights, explaining what he's posting a picture of and why?

 

[Farid]

Actually, though I appreciate Version has a particular role to play in this forum and he does it well, aren;t the rules and regulations such that he ought to add meaningful comments and insights, explaining what he's posting a picture of and why?

A few things are to note here.
It's true that people have to substantiate their posts with something other than a one-liner or an unrelated picture.

But in this case, do you really want version to explain what he meant when he posted this picture? It will probably be something related to photon mapping in RT using the 32 4GHz SPEs hidden in RSX, or something. I mean, there are things better left unsaid, that is especially true with version's posts.

 

[Rodéric]

Actually, though I appreciate Version has a particular role to play in this forum and he does it well, aren;t the rules and regulations such that he ought to add meaningful comments and insights, explaining what he's posting a picture of and why?

Looking at the picture alone, it looks like Frustum Culling with Occlusion Culling.
The Spheres would be the bounding volumes, and the grey area Occluded zone.

(It's not clear, but I suspect there's no Bounding Volume Hierarchy in this system.)

 

[Shifty Geezer]

A few things are to note here.
It's true that people have to substantiate their posts with something other than a one-liner or an unrelated picture.

But in this case, do you really want version to explain what he meant when he posted this picture? It will probably be something related to photon mapping in RT using the 32 4GHz SPEs hidden in RSX, or something. I mean, there are things better left unsaid, that is especially true with version's posts.

I did think version is awarded special dispensation, seeing as he does what he does so well

 

[Titanio]

Its also good to note that EA was not talking about the RSX but rather the 7800GTX they had in the development kits at the time. That article about MoH on the PS3 and the fill-rate bound stuff was written many months ago and with RSX's not coming out in development kits until DEcember it is impossible for them to have had it by then.

I'm not sure if the final chip would change that (fillrate) much, barring more framebuffer cache or something like that, perhaps..

It's not necessarily something that needs improvement, though. There has to be a bound, unless you have an infinitely powerful system It's really only a problem if it caps performance at an unacceptably slow framerate (and there's no indication that that's the case here).

 

[Guilty Bystander]

I'm not sure if the final chip would change that (fillrate) much, barring more framebuffer cache or something like that, perhaps..

Doesn't really matter though as the RSX still would have more than twice the pixel fillrate of that of the Xenos if we would think of the RSX as a GTX512 Core.

RSX:
16 ROPs x 550MHz = 8,8GPixel/s

Xenos:

8 ROPs x 500MHz = 4,0GPixel/s

 

[ihamoitc2005]

Fill-rate

Doesn't really matter though as the RSX still would have more than twice the pixel fillrate of that of the Xenos if we would think of the RSX as a GTX512 Core.

RSX:
16 ROPs x 550MHz = 8,8GPixel/s

Xenos:

8 ROPs x 500MHz = 4,0GPixel/s

If he is talking of pixel-shader output then difference is not so much, only 13.2Gpixel for RSX and up to 12Gpixel for Xenos (if CPU vertex shader). I think it is no problem for developers to get max shader performance from consoles. Actually many developers have said it is not problem.

He is not clear so we can only make guess but I feel EA person was talking of pixel-shader "fill-rate" not ROP fill-rate because I feel this is larger factor for limit for game with very high resolution and pixel effects. But this is only my guess. If he is talking about only ROP output it is surprising to me but maybe we can know more with next interview.

 

[Dave Baumann]

Doesn't really matter though as the RSX still would have more than twice the pixel fillrate of that of the Xenos if we would think of the RSX as a GTX512 Core.

Although I suspect that they are not talking about colour/Z fillrate, note that a 128-bit bus is saturated even before 8 pixels. All of the current 128-bit PC cards (even ATI's X1600, which has about the same graphics bandwidth as RSX) are shipping with a maximum of 4 ROPS with 4 blends per cycle. Even if all the non-writable graphics operations were shifted over from graphics to system RAM RSX would never hit its peak fillrate if it had 16 ROP's (which is why I suspect that the fillrate will be halved).

 

[Titanio]

Doesn't really matter though as the RSX still would have more than twice the pixel fillrate of that of the Xenos if we would think of the RSX as a GTX512 Core.

RSX:
16 ROPs x 550MHz = 8,8GPixel/s

Xenos:

8 ROPs x 500MHz = 4,0GPixel/s

Who said anything about Xenos?

You have to consider more than ROPs, also - bandwidth to the framebuffer would be a big factor.

Again, though, that the bound is there in itself tells us little or nothing. It only matters if its restricting performance unacceptably.

 

[Edge]

Actually, though I appreciate Version has a particular role to play in this forum and he does it well, aren;t the rules and regulations such that he ought to add meaningful comments and insights, explaining what he's posting a picture of and why?

I think Version likes to be mysterious.

 

[phed]

A basic question. Why is the GPU hertz so limited in contrast to CPUs? I mean 4GHz vs 500MHz ... It seems like it is missing out on a lot of bandwidth here.

 

[AlNom]

A basic question. Why is the GPU hertz so limited in contrast to CPUs? I mean 4GHz vs 500MHz ... It seems like it is missing out on a lot of bandwidth here.

Look at the cooling methods.


(there are other things to consider)

 

[weaksauce]

Look at the cooling methods.


(there are other things to consider)

How about bit bus?

 

["Nerve-Damage"]

Things that I believe were true:
The first RSX (probably the final design at that time) was finished in late 2004; consisting of 300+ million transistor design.

Hence this statement in this 2004 Article:

"I have every expectation that we will be able to see final production silicon later on in the year," said Huang. "This next generation architecture took several hundred people several years to build, but this specific implementation of that architecture should take about 50 engineers. It's something we're running full throttle on."

Then something happened!!

I.e. the Xenos GPU happened!!


I think to some degree Sony/Nvidia felt the finalized RSX (2004 version) may not have been up to snuff compared to the Xenos GPU; probably figuring it was 2-3 steps behind in certain algorithm areas.

Thus, Sony having no solid (design) details of the RSX during the 2005 E3 presentation; other than that of the 2004 version (RSX) performance. In other words…if all else fails with there 2005 RSX update/redesign; Sony could fall back on there 2004 taped out version.

Anyhow, I predict RSX is no longer a 300+ million transistor design (Maybe during the Dec-2004). I believe now the RSX will eclipse the 450+ million transistor mark (with two reliable sources to back up my claim).

 

[3roxor]

Anyhow, I predict RSX is no longer a 300+ million transistor design (Maybe during the Dec-2004). I believe now the RSX will eclipse the 450+ million transistor mark (with two reliable sources to back up my claim).

How reliable are those sources and in what field do they work?

 

[Titanio]

Things that I believe were true:
The first RSX (probably the final design at that time) was finished in late 2004; consisting of 300+ million transistor design.

Hence this statement in this 2004 Article:


Then something happened!!

I.e. the Xenos GPU happened!!


I think to some degree Sony/Nvidia felt the finalized RSX (2004 version) may not have been up to snuff compared to the Xenos GPU; probably figuring it was 2-3 steps behind in certain algorithm areas.

Thus, Sony having no solid (design) details of the RSX during the 2005 E3 presentation; other than that of the 2004 version (RSX) performance. In other words…if all else fails with there 2005 RSX update/redesign; Sony could fall back on there 2004 taped out version.

Anyhow, I predict RSX is no longer a 300+ million transistor design (Maybe during the Dec-2004). I believe now the RSX will eclipse the 450+ million transistor mark (with two reliable sources to back up my claim).

Crazy post, if you'll forgive me. Huang was referring to their fiscal year which ran well into 05.

The reluctance to confirm pipeline specifics at E3 and the like is likely simply down to being too far out to precisely forecast yields (and thus how much redundancy would be applied to the chip). It's possible the given figures for shader ops/cycle etc. refer to the low end of their expectations, but that'd be about it.

 

[Brimstone]

Things that I believe were true:
The first RSX (probably the final design at that time) was finished in late 2004; consisting of 300+ million transistor design.

Hence this statement in this 2004 Article:


Then something happened!!

I.e. the Xenos GPU happened!!


I think to some degree Sony/Nvidia felt the finalized RSX (2004 version) may not have been up to snuff compared to the Xenos GPU; probably figuring it was 2-3 steps behind in certain algorithm areas.

Thus, Sony having no solid (design) details of the RSX during the 2005 E3 presentation; other than that of the 2004 version (RSX) performance. In other words…if all else fails with there 2005 RSX update/redesign; Sony could fall back on there 2004 taped out version.

Anyhow, I predict RSX is no longer a 300+ million transistor design (Maybe during the Dec-2004). I believe now the RSX will eclipse the 450+ million transistor mark (with two reliable sources to back up my claim).

I get the impression that the G70's in the developer kits were put in until the final design is completed. Just like XB360 had X800's until Xenos was finished, PS3 kits have g70's until nVidia completes it's next-gen GPU. Sony conformed the PS3 timetable to fit into what nVidia had in it's product pipeline. During the timeframe the nVidia/Sony GPU contract was announced, both companies had a fair idea of what the R500/Xenos was going to be.

 

[3roxor]

Is it confirmed that at the CES they are going to introduce the RSX?

 

[Titanio]

Is it confirmed that at the CES they are going to introduce the RSX?

No, far from it. CES isn't even on the last official timetable we had for PS3 PR. You should keep your expectations low..

 

[typoEDR]

Is it confirmed that at the CES they are going to introduce the RSX?

The only thing Sony has ever confirmed about CES was the BluRay presence. Everything else is a mystery.