In terms of rendering power what would Cell be equal to?
- Thread starter almighty
- Start date
This topic cannot be answered until someone writes a rasterizer, not a raytracer. And at the moment, the only graphical interest talked about is in the field of shooting rays. Until then it's pure pie-in-the-sky guesswork. And a lot would depend on workload too. Texture-heavy graphics Cell will stumble over. Lose the textures for calculated shader effects and it'll speed up, able to do things old GPUs can't. Very much an apples to oranges comparison without anyone even trying to measure the cirtus content of an apple.
That was kind of the point I was attempting to make.
icecold1983
Banned
how does that change what dcs gpu is capable of?
i thought mesa was being ported?
true a lot of things are worse, but its doing something even the r600 + g80 cant (honestly) do, perfect reflections with the windows
whats the strange darkspot under the bench?
hmm perhaps theyre doing AO!
also shadows from the sun tend to have hard edges ( as the sun is a long way away )
Lol, that alone is enough to ignore youre post but I give you an answer neverthless. First this is about offline renderers (if you read the whole posts) like iRT, 3Dsmax, XSI. These render programs use the CPU to render and there is no limitation to what can be done except for RAM/VRAM to store the result. OF course more detailed means more time to render a frame.
Perfect reflections is your intertpretation, I say they are to bright and detailed (no shades). But all it is reflecting is the sky 2D clouds and you mean that the r600 + g80/xbox360/PS3 can't do this in a game? -
You mean the bench shadow?
Realistic shadows have different shade tones (darker/brighter areas of the same shadows) and mostly blurred edges.
The real sun or the light source in the renderer?
Last edited by a moderator:
It changes everything since I'm comparing ONE CELL not 15.
3 million polys/frame = 90 million polys/sec at 30fps using 15 CELLs therefore each CELL is only capable of rendering 6 million polys/sec. PowerVR DC is capable of rendering 7 million polys/sec and Kyro is capable of 20 million polys/sec.
icecold1983
Banned
well u said it could render that scene, not that it could render what 1 cell could.
7million polys throughoutput divided by 2 gives 3.5m/s in scene at 2fps (although VRAM space adequate for the scene is needed).
But then other effects used in the rendering has to be taken into account...
Raytracing! How many polys a second can PowerVR DC raytrace? And how many poly's a second can Cell renderer via scanline rasterization?
Until you're comparing the same operations, you can't compare performance. Raytracing and scanline rendering are two very different tasks even though the results look similar.
Err..I don't think you can assume a linear scaling like that, particularly with a rendering technique whose linear cost might be more per-pixel than per vertex. Someone can correct me if I'm wrong, but ray tracing performance tends to depend more on resolution than anything else (not exclusively, but more).
To illustrate, another Cell RT paper shows performance figures for a couple of different scenes. Their performance with one scene that has ~850x the number of triangles of another is 3 times less, not ~850x less.
According to Barry Minor, IBM will be showing soon what can be done with a single PS3 Cell using iRT, they may even release a standalone PS3 Linux application, so maybe then we'll have a better idea.
We are all aware of that, that's why everything is speculation.
I don't think a single CELL is much better than a PowerVR DC when doing purely rasterization, regardless of what algorithm that demo is using. PowerVR DC uses a Infinite Planes algorithm (Ray Casting) as well as all of the other built-in hardware features so a single CELL would have its hands full trying to emulate it in software. That's just my opinion. Do YOU think a single CELL can fully emulate a PowerVR DC chip at full speed in software?
Last edited by a moderator:
I think we maybe need to clarify some things here.
If we're talking about mapping a GPU pipeline to Cell and wondering what kind of GPU it could match, that is one thing. I think that's what the OP was asking.
If we're talking more generally about rendering techniques, then that's a whole other ballgame. I'm not sure what you're saying there Capeta, that you think DC's PowerVR would beat a Cell regardless of the rendering technique used, or just if you tried to map its pipeline to Cell..?
edit - if you're talking about emulation, I don't know..but that's sort of an impractical question anyway unless you're interested in DC emulation. If you're talking about raytracing performance or the like, I think that's an obvious answer.
If we're talking about mapping a GPU pipeline to Cell and wondering what kind of GPU it could match, that is one thing. I think that's what the OP was asking.
If we're talking more generally about rendering techniques, then that's a whole other ballgame. I'm not sure what you're saying there Capeta, that you think DC's PowerVR would beat a Cell regardless of the rendering technique used, or just if you tried to map its pipeline to Cell..?
edit - if you're talking about emulation, I don't know..but that's sort of an impractical question anyway unless you're interested in DC emulation. If you're talking about raytracing performance or the like, I think that's an obvious answer.
Again that demo used 15 CELLs so it doesn't mean much when you're talking about ray tracing using ONE CELL. What can you do with one CELL doing raytracing at 30fps? A spinning cube?
BTW yes it is my opinion that a single CELL is about equal to a PowerVR DC chip when emulating it.
BTW yes it is my opinion that a single CELL is about equal to a PowerVR DC chip when emulating it.
Haha, no, I'm sure it could do more than that. Like I said, IBM hopes to release an iRT application that will let us see what a single PS3 can do with that software.
If you want a more immediate comparison with somewhat similar techniques, there was that julia fractals ray tracing demo (on a single cell) that beat an implementation on an overclocked 7800 GT.
I'm guessing the suggestion is not that for raytracing, DC's PowerVR would beat a Cell, but I'm getting kind of confused now.
OK, who knows. But like I say, I think the whole question of what GPU Cell would be equivalent to if you comprehensively mapped its pipeline to it is a little misguided. If you want to do rendering on Cell, you probably wouldn't just naively/blindly map that kind of pipeline in its totality, unless you needed some sort of legacy compatibility or emulation. Or at least, an application using that pipeline on Cell probably wouldn't use it as they would on a GPU..you'd want to mould your application it to take advantage of the parts of the pipeline that fly and avoid pitfalls with those parts that don't. Something like the MESA project might ultimately give you an idea of how an OpenGL-like pipeline can perform on Cell, though.
wasn't there just an article a few weeks back that said a single SPE can calculate 80 million polygons a second??? or was it vertices???
I'm a big fan of power VR and all but i don't think it can sustain its 7 million polygons/sec if it were to do ray tracing.
Emulation is the only way you're going to be able to compare performance of CELL to X rasterizer, unless you have your own custom software rasterizer that does the same things as X rasterizer. Since there is no software rasterizer running on a single cell, it all speculation.
I think that's irrelevent since I'm comparing rasterization. I'm not comparing transformation rate. I have no doubt CELL is a geometry monster.
It could sustain 7 million polys/sec doing ray casting which is similar to ray tracing.
Last edited by a moderator:
