Can this SONY patent be the PixelEngine in the PS3's GPU?

This means that we'll be seeing the polygon count over 1 billion in real time then?

 

Not necessarily.

 

No, (as performance isn't really speculated upon in the patent). But it 'could' mean we won't necesserily have to render polygons at all if we feel like it...

 

That's music to my pointy hobbit hears.

 

they're either giving an example, or possibly even talking about what to expect from the PS3 or both. hundreds of millions of polygons a second. that's under 1 billion.

as said before, even if the PS3 is capable of calculating a few billion polygons a second, when we look at what the GPU will be able to display in games, it'll be more like less than 1 billion polygons/sec. probably several hundred million polygons/sec. that would be a 10 fold increase beyond the highest polygon/sec PS2 games. (in line with what EA said recently)

now if we compared the earliest PS2 games which pushed a few million polys and jumped forward into the future to see what late-generation PS3 games were doing, we'd probably see a 100 fold increase in polygon count.

PS3 might be able to hit 1 billion polygons in game at some point. but nobody can say right now. we don't know the performance of PS3, or what will be able to be sqeezed out of it, or how much shaders and effects will impact geometry performance, etc etc.

 

Several million polys per second in game would be more like a 40+ fold increase over PS2. Average in game poly per second performance for PS2 is around 10 million in the best PS2 games (around 20 million at peak). So a 10 fold increase with PS3 will see around 100 million average in the best games.

Anyway I expect between 100-200 million pps in game next gen, no more.

 

Even then, polygon rates are highly dependant on LOD algorithms. A game like Farcry is "only" pushing on average between 1 to 6 million polygons per second (depending on the situation), as good as it looks. 100 millions would already be A LOT. Huge jump over Farcry which already looks far better than most console games this generation...

 

Just FYI, you can't really go by polys per second with FarCry since its a PC game and framerate varies wildly depending on what hardware you play it on. Better to look at polys per frame. Also I think the reason Far Cry looks so good is more about effects then polygon counts (well its a good mix but mostly texture detail and shader effects).

But yeah 100 or 200 million polys per second average in game on next gen consoles would be fantastic. I wasn't trying to say that it wouldn't be a good leap or anything. Just saying that's what I expect next gen.

 

Well i was using Farcry as an example, running at full maximum detail at an average of 30fps. Polygons per frame figures are wildly dependant on "what u're looking at", and honestly i can't really remember the two extremes. I do remember that sometimes it goes below 1 million polygons per second and never above 6 million polygons per second. Farcry does look as good as it does mainly because of the vegetation, the water and the large vistas while outside, and the very nice use of pixel shading while inside, and on the characters too...
The textures are not out of this world though. Painkiller does a much better job with textures.

Still, even keeping the same texture detail but increasing the polygon count 20X would work wonders

 

comparing PC games to console games is kinda worthless, in most cases, since PC games never almost never push a specific PC configuration (CPU, GPU, memory, bus, etc).

console games are a good comparison naturally. most console games this generation push under 10 million polys. a fews push 10-20 million polys. so 10-20 million is the max. the first generation games on next-gen consoles should easily be 100-200 million polys so that would be a 10 to 20 fold increase over the best current generation games, and roughly a 40-50 fold increase over the average currentgen game. I expect next gen consoles be capable of pushing somewhere in the middle of the hundreds of millions of polys when really pushed hard. so 400 to 600 million polys with some pixel shading and some effects. that should be achivable if nextgen consoles are tranforming billions of polygons. we'll say 2 to 3 billion raw or flat shaded polygons. remember the GSCube could do 1.2 billion raw or flat shaded, from its 97~98 GFLOPs peak. even if PS3 CPU gets only 256 GFLOPs, it'll still surpass GSCube by alot.

 

And where exactly are we storing these 100's of millions of polys?

I just recently did some rough math for a product for a next gen platform that seemed to imply I could draw more polygons in one frame than I could store in main memory, and that was using a pretty compact vertex format.

It's nice to talk about procedural generation of geometry, but much harder in practice to do in a useful way.

 

Yes, but the point of me mentioning Farcry was that it is regarded as the most beautiful game out at the moment, regardless of the platform. And thinking Farcry poly counts X 20 =
Although it must be said, Farcry's asset is in the shaders...

And again, i really think that with LOD, there won't be the need to push "hundreds of millions of polygons per frame". The screen wouldn't be able to display them anyway. 640x480x60(fps) = 18,432,000.

Of course it's not that easy, but even 100M polys are a lot. And if that's a tenth of next gen maximum capabilities, Displacement mapping would be used just to throw in some eye candy for free.

I think keeping the polygon counts "low" (as in "not in the several hundreds of millions") and focusing the power on better IQ, animation and physics will be much better than just focusing on pushing a billion polygons per second in realtime....

 

I will not argue that RAM / MEMORY will be a limiting factor in the final amount of on-screen geometry next gen consoles can display

 

Using NURBS and Subdivision Surfaces ? Would that help you in the scenario you were working on ?

 

They can be useful, but my experience is that they end up not so much saving you space as allowing you to get closer to the model.

I've said this before I seriously looked at both of these technologies for a racing game a few years ago and it was cheaper to store a very highres model (>25000 poly's) than to store the subdivision surface model or the nurbs model. You have to get really close to a 25000 poly car model for the polygons to become obvious (the model had bulbs in the headlights!).

Space is going to be the big issue, that and how you light and shade the polygons you can store.

FWIW I also think that clever use of instanced geometry and textures can alleviate the issue. i.e. procedurally tweaking them as they are drawn to provide variation at a reduced cost.

 

PEOPLE- this is obviously a video adapter to accelerate window scrolling on a "home server product". I mean, c'mon, can we get a life and get back to talking about the DC legacy somemore!

 

I'm expecting way more that 100-200 million polys. I'm looking at 1-3 bpps in real time and 10-16 bpps raw.

 

Please, do expand on this and break some myths along the way.

It seems so strange to me that a NURBS model or a Subdivision surface model can take more memory storage space than a straight 30,000 polygons model.

It is also true that if the NURBS surface uses a tons of control points, you have to store them.

The idea with NURBS and Sub-Division surfaces is to make surfaces smoother and more detailed with less memory storage cost, but I can see how modelling VERY VERY small details of lots of different parts on screen ( you tend to use much more small patches rather than one single big patch ) might take quite a big toll.

Is this what you are saying ?

Thanks for the help ERP .

 

I believe one should keep their expectations modest. not saying you are wrong but isn't it better to be conservative rather than pie in sky and be disappointed?

honestly I expect 2-3 billion polys raw and 200~500 million with heavy texturing, effects, lighting, pixel shading, etc.

 

The same place where we store 'bilions of texels' on current consoles?

I agree, people are all too quick to forget the overhead asociated with various HOS approaches. Outside the ideal shapes that don't really show up in games based on real looking stuff much, the worthwhile benefits in storage don't really come until you start zooming in really close at stuff.

Screw subdivision, we'll just define NURB primitives on this new rasterizer and do away with polygons all together.