So exactly how does the Xenos emulate a Physics Unit?

http://www.beyond3d.com/forum/showthread.php?t=24771

I think Xenos cannot support some of the data formats used in Toyshop - not sure. Additionally, Toyshop would run more efficiently if it was re-coded to use memexport instead of R2VB.

Jawed

 

Although I do wonder since Xenos has 3 by 16 pipe shaders, If they were to use one of it for physics, won't that cut it's shader power by 1/3?

 

Hahaha, you know that's a Radeon 8500 whitepaper, right?

 

Yes, if that Shader Array is devoted solely to physics. But you can use it for part of the time, and perhaps even integrate some of the physics processes with the normal rendering tasks, such as providing interim vertex positions of a transformation back to the CPU for physics processing in the middle of renderng the scene.

Like all programable hardware, the devs can choose to use 1/3rd the power, all the power, or a tiddly fraction of the power for whatever purposes they want.

 

Well thinking of the already very limited fillrate of the C1 i guess it's not a good idea to leave away even more "pixel shaders" because of physics calculations and thus reduce fillrate. That's something completely theoretical which is never gonna be used in a real game on the 360 simply because there arent enough resources.

 

Well you just keep hoping....

 

Limited fillrate by no means entails limited ALU performance. A fillrate limitation may reside in the number of framebuffer reads and writes that can be accomplished, independent of the number of ALUs available, which are the untis of primary importance to a physics engine. In addition, 48/32/even 24 ALUs at 500 MHz is quite a lot, considering Xenos only has to process fragment programs for 1280*720 pixels and is limited to 8 color writes or less per clock (w/AA of course). It seems to me that in situations where the ROPs become a limitation, because of fancy post processing, etc., there will be ALU cycles to burn on physics, etc., if the required vertex/fragment processing is not enough to mask the ROP limitation.

 

So what, principles are the same, the ability how to make fluid reality and physics effect on gpu shaders.

 

Absolute bollocks. Since when was 4Gpix/s "very limited fillrate"?

You're going to see much higher real-world fillrate on xenos with its 8 rasterizers and high-performance eDRAM than on 24/32 rasterizer high-end PC parts simply because they don't have the memory bandwidth to sustain their fillrate, particulary when alphablending.

 

So considering the Cell is strong on FP calculations, or stronger than the xenon.

What sort of effect will this have on the 360, when the PS3 is capable of doing part of its physics on the CPU and part of the physics on the GPU ?. (if that's possible)

Obviously the 360 capable of doing the same but having less FP calculations to play with.

I'm just curious if their would be much benefit in the xenos doing physics calculations and falling short in shader performance, surely any advantage the xenos had over the rsx would be minimized.

 

Geometry shaders aren't just about creating and destroying geometry. They give you the ability to run a shader on an entire primitive instead of just the vertices individually. Developers will likely come up with lots of algorithms that use geometry shaders yet don't amplify/tessellate the geometry.

 

Doing tesselation often requires access to more than just input vertices, but adjacency information of those vertices. For sure, you can hack *anything* into old shader models (there was a SIGGRAPH paper that once showed that OpenGL was sufficient for Church-Turing completeness) and GPGPU shows you can put general purpose algorithms there.

The question isn't "can it do it", the question is "is it efficient and easy to do" On that basis, I'm not sure using Xenos to accomplish what DX10 geometry shaders are attempting is necessarily the best usage for it's ALU power. For that matter, I'm not sure DX10 is sufficient to address general purpose tessellation algorithms.

For the PS3 atleast, tesselation is a ready made job for the SPEs. It also should run quite well on XeCPU.

 

http://www.beyond3d.com/forum/showpost.php?p=660248&postcount=193

 

It should be, Havok's GPU library has been developed on G70 hardware.

It's questionable what advantage there is to start with, surely. My guess is if you budgetted enough ALU power on the GPU to do something worthwhile, you'd be talking about clear disadvantages rather than "minimised advantages".

It is true that if you are not ALU-bound, you effectively have some ALU power to spare. But how much would you need to make it worth your while? How much could you do before you became ALU-bound?

It's also inevitable that, as a closed box, software will catch up more quickly with the hardware's capabilities, such that if it's not already the case, it will be reasonably quickly the case that (m)any games won't be so under-utilising the GPU's ALUs as to present a meaningful opportunity for physics. IMO. Devs will be stretching the GPU's abilities for rendering alone, over time, and I'm not sure how typical the desire to sacrifice the graphics budget for physics will be.

Of course, one could argue that the type of physics we're talking about here - "effects physics" - are directly visual, and directly impact the visual quality of the game, and thus are as worthy of GPU time as anything else. And you'll see no argument from me there...but wouldn't it be nice to do those things without trading off against rendering power for everything else...hmm...

It's a nice option to have depending on a developer's priorities. But I just wonder how the typical game will be pushing the system by the time libraries and such are available to do this, and how much of an opportunity would be typically there for physics to "steal" some GPU time, given the more 'traditional' graphical demands of future games.

 

Well only half of a regular G70 isn't it? Consider a 550MHz G70. 360 already struggles on High resolutions even without tiliing and AA. (There was a comparison of 480/720p in a computer mag of games like PDZ and CoD (current games) not long ago which acknoledges this, 480p always considerably faster than 720p).

 

Uhm doesn't the X360 always render at the same resolution internally, to then scale it to whatever you choose? That's what everyone has been saying.

 

Guess that's the same thing like "all games have AA" (M$).

 

Fillrate is more than a function of ROPs. You also have to consider bandwidth. I'm not sure what a G70's effective fillrate is.

Even if true, it wouldn't necessarily indicate an issue with fillrate. In such a scenario you're not just painting out less pixels, but shading fewer pixels (which would relate to shading/alu power).

 

1st of all there is not a single game on 360 that does not have AA. The worst offendor, is PD0, and it still looks like heaven compared to the inch long jaggies that plagued us last generation.

Secondly, there is not a single 360 game that does not use tiling, it's impossible not to. They just don't do it well.

Thirdly, I have both COD2 and PD0 and anyone who says that either game is faster at 480p is completely exagerating, PD0 doesn't even have slowdown, and COD2 run pretty much buttery smooth the entire time. Does it run marginally better at 480p? Maybe, but it's nowhere near "considerably faster"

 

Wasn't it noted prior to launch that MS would accept things like motion blur or DOF as 'AA'? Very close to launch PD0 did not have any explicit FSAA, and this was confirmed by Rare themselves (who went on to talk about the motion blur etc. as helping out here). Did anything change before the game was launched?

I'm not sure that's true. The only reason PGR3 would be using '600p' would be to avoid tiling, otherwise they may as well have used 720p.