Looking good. Not much of a surprise really since Sony dev kits have been using 6800U and 7800GTX GPUs as a base design to go forward on development. Not to mention a couple devs straight out saying, "RSX is a modified NV40/G70". ATI in more than one interview has called RSX a G70 derivative. Which of course all goes back to the clock-performance (shader ops, flops per cycle) slides at E3 (which indicate a G70 derivative at 550MHz) and the NV stock holder statements that RSX would be a G70 derivative and "therefore will cost very little in new investment". I don't expect NV to lie to their stockholders about the R&D costs of such a deal considering it could mean prison time.

That said, lets look at some of the numbers.Technically true... but there is more to it. RSX has a 128bit interface to the 256MB GDDR3 pool (22.4GB/s). But it ALSO can access R across the FlexIO which offers 25.6GB/s in bandwidth just to R memory bandwidth.

While CPU intensive games are going to need 10-20GB/s of that bandwidth, RSX will have more than 22GB/s of bandwidth. I am guessing a lot of initial games will be more eye candy (as devs figure out CELL) and so a good portions of the 48GB/s will be available to RSX. I think this namely because floatingpoint type math can work with small applets and produce a lot of information separate from main memory accesses in a lot of situations (which will be necessary with the SPEs) and because a top end CPU like an AMDFX-57 does pretty solid on top end PC games with only 6.4GB/s of memory bandwidth.GDDR3 22.4GB/s
R 25.6GB/s
-------------------
48GB/sThat puts it at 12.5B for RSX. This will take a big hit at higher resolutions with AA, but I don't think realistically many games will be going for anything other than 720p anyhow. Should be more than adequate.I will have to check, but I think that is the setup engine (e.g. Xenos can setup 500M vertices/triangles a second--one per clock--or tesselate 250M per second--one per two clocks; yet the shaders themselves can transform like 6B+ triangles).

The reality check is this: Triangle Meshes are HUGE. They consume a lot of memory bandwidth and memory footprint (unless you procedurally generate them). This is why normal mapping is very popular. The detail of a 20MB mesh can be stored in a couple MB normal map and also be processed by pixel shaders. Lower detail meshes also means faster shadows and less work in transforming them for animation. This is why games like D3 and UE3 are Normal Map heavy.

So the number being huge is nice because it means it wont be a bottleneck; on the other side it does not tell us too much. To put it into perspective of a 720p screen:

720p = 921,600 pixels
720p * 60fps = 55,296,000 pixels
1.45B / (720p * 60fps) = 26.2

What this means is that PER FRAME, RSX theoretically can have 26.2 triangles *per pixel*. (Xenos, with the 500M setup limit, is 'limited' to 9 triangles per pixel at 720p).

Baically theoretically these things can setup more pixels than you could ever see. Practically there are many other bottlenecks in the systems that make these numbers irrelevant.The E3 slides always indicated G70-like shader pipeline @ 550MHz. Of course this is NOT an official announcement of RSX, but then again everything lines up right.

Overall looking hot.

As for G71, a nice ~20% core clock boost over the 7800GTX 512MB. On the down side no improvement on the memory. Unless they did some major tweaking to their SM3.0 implimentations I am not sure this is good enough to best the X1900XT. Why? Because in shader heavy games the X1800XT already edges the G70@550MHz in FEAR, CoD2, and ties in BF2 and the X1900XT walks away. A 20% boost in core performance is GREAT, but ATI took a gamble and increased their pixel shader performance peak to 3x the X1800XT. Of course most current games don't need that much extra pixel shader performance and the G71 is gonna get a boost to ROPs and TMUs as well to shaders, but future looking as shaders become more important I think ATI made the better bet. I expect a splitting of a lot of benchmarks, with G71 taking a lot of games that use heavy texturing and R580 taking the more shader heavy games. Choice is good :D

Of course it is all moot as $500+ GPUs are not the mainstay of most consumers. All bragging rights.

Whoa! Slow down guys! I say the first 3 posts when I started, wrote, posted and now there is like a dozen! Slow down for me!Vertex = Triangle, at least for this type of math. In game usually there are 1.2-1.5 verticies per triangle.

(Just think of it this way: Every time yyou add a new vertex you can link it to a number of others which produces more and more triangles).

Heh, after skimming the rest of the thread it seems my 6B max for all 48ALUs on Xenos was right (from memory). Of course as Jolli is saying we will NEVER, EVER see that. Ever. (Period).

You are confusing the ability of the shaders with the setup.

Xenos can *setup* 500M triangles--max.

In theory, if the above limitation was not set, it could transform 6B triangles if every resource was dedicated to such and no other bottlenecks were hit.

*But* vertex shaders do more than just transform verticies. e.g. some shadowing techniques and displacement maps use vertex shaders to produce their effects.

If all Vertex Shaders did was transform meshes then we could still have a Fixed Function T&L setup.

You are right about D3D compression from Xenon (believe it is 2 fold compression making the 10.8GB/s one way effectively 21.6GB/s) but remember, at its best, Xenon is going to be able to do 1/3 the work Xenos could do--and that is theoretical (no way a CPU is going to perform as well as a finely tuned multithreaded GPU pipeline). Anyhow, if Xenon pushes 2B triangles to Xenos it cannot handle them--because, again, it is setup limited to 500M/s.

Which of course is more than we need, even with significant overdraw.

So, as I said yesterday, lets not get hung up on irrelevant numbers. Lets just toss out shader ops and setup limits or peak vertex transformations because none of them tell us anything about the power of a GPU, its bottlenecks, or how it will apply to real games.

Yeah we do, it is 500M vertices. One vertex per clock.

Think of a octagon (6 sided poly) with a dot in the middle. Draw a line from every corner of the octagon to the middle dot. The 6 points of the of the octagon + the 1 point in the middle can form 6 triangles. 7 vertices, 6 triangles.

In larger meshes you can make it so 1 triangle = 1 vertex.

And thus in GPU nomenclature 1 triangle is the same as 1 vertex in regards to such. And besides this it was confirmed numerous times on B3D.Both. Depends what part you are asking about. What part/numbers is confusing you?

I got that number from a developer (ERP). It is variable and depends on the mesh size and design.Most 3D apps, especially ones with mesh optimization features, will output this information for you Not sure I understood the first part about larger objects have fewer vertices? Larger how?Setup = the maximum number of vertices that can be setup for the pipeline. Since everything travels through the pipeline you can never have more than the GPU can setup. The setup number is basically => How many can the GPU "catch" and use.

GPUs, ever since the advents of programmable vertex shaders, have been able to transform more vertices than they could setup. This is because the VS are expected to do more than just push the mesh through.

If you check my posts the other day to the Displacement Maps you can get an idea of WHAT the power of vertex shaders can be used for. This is why they have more power than the setup limit--basically, if they were the same (or less) you could do NO vertex shader effects less setup the triangles.

I think VS are often misunderstood. This was the state of offairs before and after VS were first introduced waaaay back when:http://www.gamedev.net/columns/hardcore/dxshader1/...

Vertex Shaders are even MORE capable now; and while they are frequently ignored over PS due to some limitations (now disappearing--like being able to fetch a texture) they are becoming more and more important.

Hopefully that helps explay why Vertex Transformation numbers > Vertex Setup numbers. Basically the VS units have to do a lot more than just setup the triangles--but of course the GPU can NEVER output more triangles than the setup limit.

Which is irrelevant due to the target resolution (720p) and the huge size of a 1.45B triangle mesh!

Of course note--that is like saying the Radeon X850XT is almost as fast as Xenos since it can do 43B shops.

These are all just "limits". The reason some GPUs have really high numbers that look out of place is because they are derivative models. e.g. The X700 has a high frequency which inflates the setup rate, but since that high of a number is completely unusable you will see GPU makers cut down on the silicon to setup up triangles as the frequencies go up because there is no reason to have a GPU create 10B triangles.

Die space is precious so no point wasting it on stuff never needed.Just to bring it up... we can remove Fillrate as a metric as well. In almost every case in modern games, minus reaaaaaaly high resolutions with a lot of AA, you wont be fillrate limited.

And even then it is misleading. "Oh no, Xenos only does 4GP!" Yeah, but (a) it does 16 gigasamples when AA is enabled (i.e. no penalty for 4xAA) and (b) with the eDRAM this is a "no stop" real, goodness to life number. A typical GPU, even with 16Gigapixel fillrate, is NEVER in its wildest dreams even get close to that number in a real game.