You are right, RSX is actually limited to more in the range of peak:

GDDR3 Read/Write: 20.8GB/s
R Read/Write: 15.5GB/s (read) / 10.6GB/s (write)

So RSX is actually lower than what Has is saying. And this doesn't even begin to account for the penalty of using R for texturing or inability to use it for framebuffers (i.e. pretty much limited to texturing, with penalty, or storage).Oh where to start... I know I have explained this before, so I will be brief and you can go back to ignoring it: The split in the dies is intentional based on performance needs. There is a natural break between the ROPs and the Logic.

The 32GB/s Parent Die <=> Daughter connection was designed to meet the peak needs of communications between the Logic and ROPs. An exce is MRT, where you would output form the ROPs to system memory and then re-use the render target. Of course in this situation you have 32GB/s to transfer such (60% more bandwidth than RSX has for such) -- and you still haven't touched the 256GB/s of eDRAM bandwidth for fillrate (MSAA, Z, Alpha, Color).

Yes, there is "only" so much that can be done in eDRAM... like draw your image? It isn't a general memory pool, but it is a design choice: your framebuffers are your largest bandwidth client (hence 256GB/s of bandwidth). There is a reason that RSX has a fillrate of ~2GSamples with 4xMSAA and Xenos has a fillrate of 16GSamples. Could it be that RSX is incapable of supporting the fillrate (hint: the GDDR3 doesn't have enough bandwidth in the PS3 to support peak fillrate on RSX, let alone peak fillrate + texturing and everything else).Yes, it is naive because while it is well known that the 256GB/s in the 360 is limited to framebuffers, people keep pretending like RSX bandwidth is all merry when in fact R has significant gotchas in performance and accessiblity and GDDR3 prevents RSX from hitting peak performance rates.Is animation a float limited problem or maybe a branch limited problem? Is it something that can be done through a 128KB local store or is it better suited for cache?

Not that it matters: Every developer I have talked to says animation is an *artist* limited problem. And not every problem is equal in regards to hardware. A solution on one platform is not good on another. e.g. Cloth animation and Inverse Kinematics may not have proportional solutions.Ok, answer me this (seeing as my post, last on page, in the other thread was severely ignored):

How does a Radeon X1600 outperform an GF 7900 in Parallax Occlusion Mapping? (POM, sometimes called virtual displacement mapping, is a common 'next gen' effect that gives a pronounced impression of geometry from many angles; Kameo uses this a bit and Motorstorm uses it for the mud deformation).


Ps- as I have mentioned before, both consoles have strengths and weaknesses, and will each have technical game designs that really show them off. But the "RSX 2x the shader ops / Cell 2x Float" stuff is really misleading and am surprised people still buy into it. NV30 was theoretically superior to R300 in a number of checkboxes and peaks, but in performant features we all know that all the PR hype checkbox-PR numbers got cut down by a finer look at what they actually MEANT for the architecture in said environments... oh well.

RSX has some strengths, as I have mentioned before. It has 50% more peak texture fillrate (although the dependancy upon Pixel Shader ALUs skus a direct comparison to Xenos which has decoupled TMUs). It is a more traditional design which should allow PC porting quite easily on the renderer side. It is also easier to troubleshoot at this time because of the discreet PS/VS. You have solid numbers of shader execution times whereas with Xenos devs are currently limited to shader resource needs (i.e. the threaded nature of Xenos has concerete ALU resource, like RSX, but since it is much more threaded the total execution time is not available, at least not yet). NV has done a lot of work with OpenGL and the TWIWMTBP program from NV has broad support in the PC industry. The lack of eDRAM also means no messing around with tiling, and a single pool of 256MB GDDR3 dedicated to the GPU allows some flexibility within the bandwidth budget (Xenos has to share the 512MB GDDR3 pool). RSX also has good branching in Vertex Shaders. Overall RSX is an excellent SM2.0 GPU with SM3.0 features available (albeit core SM3.0 features like dynamic branching in Pixel Shaders and Vertex Texturing are very poor, which is not surprising since RSX is based on the NV40 family).

Show me a quote.

What I see is "theoretical limit" within the context specifically of "pixel rendering to memory" i.e. graphics related bandwidth. You jumped up and down about Xenos number, when it is a peak [real] graphical limit for the framebuffers, yet never a peep about the other side.

Just as easy to say, "So you say Jolli yet people around here still throw the ~50GB/sec around as if it's the real bandwidth of the whole RSX GPU".

You cannot have it both ways.? Strike 2. First was calling others niave, now this.

Look, we already know that you have a decided preference based on PR (shops, gflops with disregard for architecture, little consideration of featureset, and so forth).

But if you keep posting "one-up" style comments, and I take the time to respond and your effort in reply is insinuitory and inflammitory replys then you will take a vacation. Either stick on topic without name calling (which you have been banned for before) or stop posting.Did you see Lair? Supposedly the past close up shot was Realtime AND the thousands of high detailed dynamic shadowed dragons on screen vid was also realtime. Yet the finished result using the same art looks HORRIBLE.

Did you see MGS4? The shadowing/lighting was poor and the texturing is very low quality.

Did you see Motorstorm? Like GT:HD, which cut down many of the particle effects from the PS2, is lacking a significant amount of particle effects.

Anyhow, you are framing the issue as a "right now" question instead of "is there an issue, and will it become more pronounced as games evolve graphically?" There is a reason why 7900/X1900 class GPUs have 50GB/s of framebuffer bandwidth. And even then GPU tests show they still can become memory limited.O'Realy?

I have read more than 1 account about Gears of War being the absolutely best looking game graphically coming out in 2006 and arguably 2007.

I don't even need to begin touching Bioshock, Mass Effect, and so forth (games again flooring people and big winners at E3).

You say you are trying to be "realistic" -- from everything I have read, realistic is: Both consoles have some great looking titles.VT3 runs at 13x7 on a 6800GS. Hitting 1080p with a GPU that in averages games is ~2x faster is not a surprise--especially with Sega's world renowned AM2 working on it. Lair... not a graphical benchmark. RR7 is fairly straight forward, ran fine on the 360 so extra dev time isn't surprising to get great results.

Anyhow, your question is like me asking how MotoGP can do what it can at 1280x1024 with MSAA.

Btw, And you STILL refuse to answer my question about POM on G70 vs. the X1600. I answer a lot of your questions.

Maybe you will go back 10 months and answer my question how 78B shader ops is a metric of performance (when PS3 devs say it isn't) and yet 2x the shops produces nearly identical Floats?