News Playstation 3 GPU is still in development

[Tim S]

http://www.bit-tech.net/news/2005/05...n_development/

[Darkedge]

with the sony conference you have to remember it was a masterpiece of PR. Very little was in realtime, they haven't finalised dev kits yet. Most were renders but looking at the videos on the net for it it seems that quite a few were possibly runing on Ultras in SLI mode - there is a noticeable frame misysnc towards the bottom of the screen. This could be the encoding but have seen it in several vids.
Other thing is the spec announced was only 24 pixel pipelines - lowr than ATi's 32... but again Nvidia is damn good at PR too, and can change this.

[Tim S]

A larger number of pipelines and a higher clock speed doesn't necessarily mean that the GPU is automatically faster. It all depends on how efficient each of those pipelines are, and how much work each one can do per clock cycle.

[Reaper_Unreal]

According to interviews with Sony, all of those demos were running in realtime, some of them were scripted, and the Killzone 2 movie was actually a movie, but everything was rendered in real time on what they called spec'd hardware. What I imagine they mean is that it was all running on their development systems.

If you ask me though, once the RSX generation of GPUs come out, I don't really see how much further they can push the current rasterization-style GPUs. Rasterization is starting to show its age. With scenes growing increasingly more complex, and with all of these crazy algorithms being used to do the effects. If you ask me, it's high time that GPU manufacturers started thinking about raytracing. First off, due to the way that raytracing works, the rendering time will scale logarithmically with scene complexity compared to linearly with rasterization. This is why we can see billion-polygon raytraced scenes without too much trouble. There's also some things that raytracing can do trivially that are quite difficult with rasterization. Shadows, refraction, reflection, implicit surfaces, non-polygonal geometry, voxel landscapes, it's all quite possible. Not to mention the fact that certain algorithms, such as relief mapping and (sort of) parallax mapping use a simplified form of raytracing and they're running on current GPUs.

Don't get me wrong, I'll buy a PS3 and enjoy it just as much as everyone else, and I'll still buy a next-next-gen ATI or NVIDIA video card, but I think that it's getting close to the time when raytracing needs to make an appearance.

[Bindibadgi]

I have to absoultely agree and i dont know why they are so keen on traditional computational rasterisation algorithums versus reallife simulation of photons through raytracing. They've already got raytracing chips developed. Maybe we'll see Raytracing and PPUs integrated into the generation AFTER these?? When they finally realise that the only way to get more is to get "real" and highly specialised.

I spose it's chicken and egg - noone will produce raytracing chips cause noone is coding for them and noone is coding for them cause there's nothing out there thatll run it.

Also, if you move away from traditional methods you loose back compatibility or make expensive dual purpose raster/ray chips. And considering ATI/NV invest heavily and know the traditional methods better along with the "short" timescales of console development maybe it's too soon.

[Bowen]

Originally Posted by Darkedge

with the sony conference you have to remember it was a masterpiece of PR. Very little was in realtime, they haven't finalised dev kits yet. Most were renders but looking at the videos on the net for it it seems that quite a few were possibly runing on Ultras in SLI mode - there is a noticeable frame misysnc towards the bottom of the screen. This could be the encoding but have seen it in several vids.
Other thing is the spec announced was only 24 pixel pipelines - lowr than ATi's 32... but again Nvidia is damn good at PR too, and can change this.

I'm sorry but I completely disagree with you. It's not that I think Sony is incapable of lying but why would they? When the PS3 is launched then the truth will come out and they'll be called out on it. I'm sure if they were lying that would come out much sooner than the actual launch date.

As amazing as the Killzone 2 demo looked why is it that hard to believe that is wasn't real time? It was running SEVEN 3.2GHz Cell CPU's...I mean come on. Even if it was running one GeForce 6800 Ultra, that would probably be enough combined with all those procs! If there was only one CPU and a 6800 then I might truly question it but with that hardware combined with the fact that Sony would really hear it from everyone if they lied about it tells me that it was indeed done in real time!

Just look at the game Gears of War from the Xbox 360...it looks pretty darn good too and it has only 3 of those IBM procs along with the new ATI card. Of course time will be the only way to truly tell if Sony was telling the truth but I think they are.

[Da Dego]

Originally Posted by Bindibadgi

Also, if you move away from traditional methods you loose back compatibility or make expensive dual purpose raster/ray chips. And considering ATI/NV invest heavily and know the traditional methods better along with the "short" timescales of console development maybe it's too soon.

Perhaps what we could see is a better use of that 2nd PCIe slot. One would contain an older vid card to run old algorithms, the other a raytrace one. That would allow backwards compatibility for a while, as well as a use for the old hardware until it became possible to put both on one board without cost, size, or heat problems (which would probably be in about 1-2 years after the changeover).

I mean, let's face it...SLI is really a bit of a dumb concept. Inefficient, overly heat producing. And if you instead developed a card based around raytrace, it would be a universal blessing (3d modellers, graphic artists, etc would also be snapping them up!) and therefore get a bigger sale than something like the enthusiast cards we see today. A raytrace card with 256mb of memory would be FAR more efficient, you wouldn't need all these stupid shader models and pixel transformations. It would simply be "object? Refractivity? Reflectivity? Opacity? Light source? Done." It would also work MUCH better with the PPUs that are coming out, because those would be able to send real coordinates to the GPU which would be able to raytrace the light sources "on the fly." It would be a HUGE step in realism, and the whole thing would probably run better (and be a little cheaper) than things now because that whole post-processing angle would be able to be eliminated.

EDIT: SERIOUSLY, in thinking about this...with the PPUs coming out, this would be a HUGE step in efficiency and realism...do you realize HOW MANY FEWER calculations would need to be done to raytrace sources, which would automatically create shadows and particle effects, etc? It would be SO much simpler than the bloat of today's chip designs...Even better, there are a lower number of equations altogether. You would essentially just need to constantly compute the falloff of light and apply that figure to the textures. Something that a fairly dedicated chip design should be able to easily do. Our slowdowns would then only be due to the number of light sources needing to be compiled instead of number of polygons. And programming for that would become a snap...They'd release raytrace engines like physics engines now...Oh, the efficiency...

[Reaper_Unreal]

Originally Posted by Da Dego

Perhaps what we could see is a better use of that 2nd PCIe slot. One would contain an older vid card to run old algorithms, the other a raytrace one. That would allow backwards compatibility for a while, as well as a use for the old hardware until it became possible to put both on one board without cost, size, or heat problems (which would probably be in about 1-2 years after the changeover).

I mean, let's face it...SLI is really a bit of a dumb concept. Inefficient, overly heat producing. And if you instead developed a card based around raytrace, it would be a universal blessing (3d modellers, graphic artists, etc would also be snapping them up!) and therefore get a bigger sale than something like the enthusiast cards we see today. A raytrace card with 256mb of memory would be FAR more efficient, you wouldn't need all these stupid shader models and pixel transformations. It would simply be "object? Refractivity? Reflectivity? Opacity? Light source? Done." It would also work MUCH better with the PPUs that are coming out, because those would be able to send real coordinates to the GPU which would be able to raytrace the light sources "on the fly." It would be a HUGE step in realism, and the whole thing would probably run better (and be a little cheaper) than things now because that whole post-processing angle would be able to be eliminated.

EDIT: SERIOUSLY, in thinking about this...with the PPUs coming out, this would be a HUGE step in efficiency and realism...do you realize HOW MANY FEWER calculations would need to be done to raytrace sources, which would automatically create shadows and particle effects, etc? It would be SO much simpler than the bloat of today's chip designs...Even better, there are a lower number of equations altogether. You would essentially just need to constantly compute the falloff of light and apply that figure to the textures. Something that a fairly dedicated chip design should be able to easily do. Our slowdowns would then only be due to the number of light sources needing to be compiled instead of number of polygons. And programming for that would become a snap...They'd release raytrace engines like physics engines now...Oh, the efficiency...

Of course one of my favourite parts about this is that the processors needed to do this aren't really specialized at all. All that's needed is a fairly efficient vector processor, something like what's in the NEC Earth Simulator, or, dare I say it, the Cell processor. We (programmers) wouldn't be struggling to make some simple volumetric effects, they'd be something trivial. Perfect shadows: check, massive polygon count: check, complex effects done easily: check. I've dabbled around in software rendering, and the big thing I've noticed is that not only is raytracing MUCH easier to write, it's much easier to do camera effects like zoom and parallax. Camera movement becomes just that, moving a camera, as opposed to moving everything BUT the camera. Lighting becomes a snap, and anti-aliasing just means that I shoot more rays.

The BEST thing about raytracing though is that it is trivially parallel. This means that it could take FULL advantage of the new dual-core processors, and all of those SSE1/2/3 instructions. This means that you can have lots of cheap processors doing the same job as a single big and expensive processor.

To be fair however, rasterization does have it's good points. There's already plenty of hardware for it. There's oodles and oodles of algorithms for doing NEARLY anything, albeit sometimes slowly. The most important thing though, is that it's a faster algorithm when there isn't much geometry. That is why it's been king of the games industry for so long.

Now if only people would do something about this and release a raytracing card. That 2xPCIe idea is brilliant. There's also a piece of software called OpenRT that does what OpenGL does, but raytraces it. I don't see why there couldn't be some sort of wrapper to port it all to raytracing.

[Da Dego]

Originally Posted by Reaper_Unreal

To be fair however, rasterization does have it's good points. There's already plenty of hardware for it. There's oodles and oodles of algorithms for doing NEARLY anything, albeit sometimes slowly. The most important thing though, is that it's a faster algorithm when there isn't much geometry. That is why it's been king of the games industry for so long.

Now if only people would do something about this and release a raytracing card. That 2xPCIe idea is brilliant. There's also a piece of software called OpenRT that does what OpenGL does, but raytraces it. I don't see why there couldn't be some sort of wrapper to port it all to raytracing.

Rasterization does have its good points, I will acknowledge. But those points are largely because we never made the leap to raytracing (the technology for truly off-loading heavy computations was just not there). As you said, algorithms for doing nearly anything...that *nearly* encompasses many things that raytrace could do more easily and with better results.

As for someone putting out a raytrace card, I'd love to see it. There's no reason it couldn't work in harmony with a normal raster card, which would allow people to develop new games for them while old games still function. Problem is, it would probably reduce the money in the manufacturers' pockets. The technology would be WAY simpler, so it would make a lower-priced card. Since most markups are based on percentage of base cost, they make much more off of a $300 card than a $100 card. Also, upgrades would be harder to sell, as the only need for them would be an increased number of light sources (you can't exactly make them more complex, light is what it is). Overall, they'd create a technology that would be hard to beat for a few years, something that would kill their sales.

Just like oil companies and hydrogen fuel sources, I would bet no one is going to move towards this until they can't make a dime more on current hardware.

[BigHype]

I think RSX = Two core G70/90nm