super-high poly 3d chip

[dpopiz]

I've wondered if it may be possible to, instead of making 3d chips that can process lighting and pixel shading and things like that that make the scene look better, make a 3d chip that simply focuses on processing scenes with no fancy effects, maye just very simple flat gourad shading or something but a HUGE poly number. I've always wished we could do things like 3d rollercoaster tycoon and other such strategy games where it would be really nice to have the scene represented by some very simple 3d. I don't know how it would work exactly, with the huge amount of geometry data that would need to be transfered, maybe some way of representing objects in a simpler fashion with lots of compression to decrease the volume of data going to the graphics card and the processing that would have to be done by the cpu.
what do you think about that?

 

[Sahakiel]

I think it would be rather pointless considering the way to represent 3d objects on a 2d surface (your monitor) is the differences in lighting on various points.

 

[FishTankX]

They exist..

They use them for working on wireframe grahpics, and most of the time they're not cards, they're computers (SGI 2CPU workstation ring a bell? One heck of a powerhouse) running software T&L with a simple rasterizer card (Imagine Voodoo2 on sterroroids). Oh, by the way, your theory on passing a huge amount of polygons, that's what AGP8X is suposed to do, my friend. The only difference though, between an SGI workstation and what you're talking about iswhen they're doing the actual frame generation part (I.E. Rendering) they don't just use gourad shading, they use like, 16 lights! So it's like, huge amount of polygons, even more insane lighting, and like.. no textures. At all. Then there are things that.. need effects, or run much more efficently if they use effects. Fur, water ripples, etc.. vertex shaders all do them more effectivley than a plane old T&L driven interface. But the Geforce4 can already process some 75 million triangles a second.. and it's real life triangle rate might be around 10 million with all the lighting and Sh!t. Why not make a scene with 500K polys and less lights? Current graphics cards are capable of what you are describing, Painkiller. It's just that you ahve to turn off lighting and do more transforming.

 

[dpopiz]

really? you sure a gf4 could do a giant rollercoaster tycoon map in realtime 3d?! I don't think so... I think you'd need > 10million polys for that

 

[Agent004]

I thought GF4 can handle just about 10 M polies

 

[FishTankX]

It could handle it, just at a relativley sad 10FPS. You see, the rollercoaster scene in Rollercoaster tycoon was rendered one frame at a time and outputted to full motion video file I.E. MPG etc.. it was drawn up in wireframe and deliberatley made to look silly through the use of simple shading etc..

 

[dpopiz]

hmmm I dunno I thought that kind of performance for what I'm talking about was still a ways away <---edit: refering to agent

anyway, also I think the main problem would be that the cpu and memory would be bogged down with too much data -- it seems to me that there would have to be some way of more simply describing the scene to the graphics card, so that huge amount of processing and volume of data could be confined to it, maybe for something like the rollercoaster tycoon example at least, just the models (track sections, cars) could be sent to the card along with coordinates for placing them so that the bulk of processing and enourmous amount of data could stay on the 3d chip and its memory system instead of clogging up the main memory and also cpu

 

[Shalmanese]

what about something like the PS2, wasnt that designed to push massive amounts of polys (2560 bit pipe)

 

[BenSkywalker]

PainKiller-

"I've wondered if it may be possible to, instead of making 3d chips that can process lighting and pixel shading and things like that that make the scene look better, make a 3d chip that simply focuses on processing scenes with no fancy effects, maye just very simple flat gourad shading or something but a HUGE poly number."

They have had dedicated T&L chips, seperate from rasterizers, for many years now. With today's fabrication technology it just doesn't make much sense anymore(you can fit the transistors on die to handle it all wih one chip).

I've always wished we could do things like 3d rollercoaster tycoon and other such strategy games where it would be really nice to have the scene represented by some very simple 3d.

You could handle that today, actually wouldn't even be that difficult. Trick is to use a geometric LOD system and 'dumb down' the geometry the further away you move, pull in closer and the geometry tesselates down to more detailed shapes.

"maybe some way of representing objects in a simpler fashion with lots of compression to decrease the volume of data going to the graphics card and the processing that would have to be done by the cpu."

HOS- Higher Order Surfaces. Everything GeForce3 and newer supports it one way or another, although you are likely thinking of more sophisticated levels then what is currently available.

"really? you sure a gf4 could do a giant rollercoaster tycoon map in realtime 3d?! I don't think so... I think you'd need > 10million polys for that

At 10Million polys per frame you are talking much better then CGI movie quality most of the time. I think you'd find one tenth of that would be plenty enough, particularly if you were using a geometric LOD system where even a million polys per frame would be a bit much(you could pull it off with much less).

anyway, also I think the main problem would be that the cpu and memory would be bogged down with too much data

I alreday mentioned HOS, to expand on that a bit currently it works by doing things like defining a point for the center and a radius to determine volume and then tesselates. Even if we skip that though, AGP 1X is enough to handle about 10Million vertices per second, and that is without buffering vertex data to on board RAM. Compared to texture data, geometry is very low on memory and bandwith requirements.


FishTanx-

"They use them for working on wireframe grahpics, and most of the time they're not cards, they're computers (SGI 2CPU workstation ring a bell? One heck of a powerhouse) running software T&L with a simple rasterizer card (Imagine Voodoo2 on sterroroids)."

Actually, the run of the mill SGI dual CPU workstation uses a hardware T&L based graphics board. The real nast poly pushers are the IR3 stations which push over 100CPUs and base in the six figure range(that is equipped with only a handful of processors).

Shalmanese

"what about something like the PS2, wasnt that designed to push massive amounts of polys (2560 bit pipe)"

Actually, the XBox can handle more polys then the PS2. The PS2 does have massive data bandwith between certain components, and it can handle a lot of polys, but the XBox can actually best it fairly soundly.

 

[FishTankX]

You know, Ben Skywalker, I was thinking of the infinte reality engine (Their graphics super computers that go up to 512 CPU's) when I meant software T&L and one heck of a power house. It just came out wrong.

And Painkiller, what you would need for such a super 3D chip would be quad T&L blocks running at 700 some MHZ and 20GB/s of bandwidth. Comming to a computer near you, in the year 2005 when textures start looking dated.

 

[Nefrodite]

what about something like the PS2, wasnt that designed to push massive amounts of polys (2560 bit pipe)


if you look at anands take on the ps2 you'll realize that sonys marketing division really likes dishing out the bull

 

[GoldMember]

Here's my stupid 2cents.. But why don't we make graphic cards with several chips? One does lighting.. one does poly's.. one does triangles and what not.. see what I mean? That way you could really great graphics.. really good FPS... with out straining your card.. smooth play. Might be expensive.. sure but what the hell.. we are already spending 300 to 400 on video cards.. why not another 100 to get something that is way better? just an idea.. I think voodoo was on that track.. but they were doing more or less load balancing between 2 chips. I wish voodoo was still around.. I'd probably would of bought the voodoo 6 or maybe even 7 by now.. over my gf2 ultra.

 

[br0wn]

<< I don't know how it would work exactly, with the huge amount of geometry data that would need to be transfered, maybe some way of representing objects in a simpler fashion with lots of compression to decrease the volume of data going to the graphics card and the processing that would have to be done by the cpu. >>



This is very interesting.

In my opinion, the goal to handle super high polygon counts (like in the order of several BILLION triangles per model) should be left to software and should not be done by creating higher bandwidth chip.

In fact, much research in recent years have been in visualization or walkthrough for model with massive amount of polygons, such as city and terrain walkthrough applications. Our research group is trying to aim for visualization of real city models (around 1 BILLION triangles) and still achieving real-time frame rates (30 frames/sec) by the end of this year. Imagine if we let hardware to handle this amount (30 x 10^9 triangles per second), no such hardware will exist, not for another 10 years.

So our approach is to combine two techniques:
1. Level of Detail (LOD)
Basically, objects that are far away can be render with fewer triangles.
2. Visibility computation
This means that only objects that are visible (or in the scene) should be sent to the graphics pipeline.
There are much research that perform either LOD or visibility computation, but very few that try to combine these two techniques into one framework, and as far as we know we are the first that has a very ambitious goal to handle a model around 1 BILLION triangles in real-time.

my 2 cents

To BenSkywalker, nice to see you still contributing here. I enjoy our last conversation about the future direction of computer graphics.

 

[GoldMember]

br0wn.. talking about only rendering that which is visible on the screen has already been done in the Kyro chipsets by ST Micro. Power VR.

They can take a very high poly count and still play it with good FPS.. because the card is only rendering what you see.. unlike the Nvidia line.. accept for the GF3 which does this a little bit.. and the GF4 which probably does it more.. The Nvidia line up to I believe the GF3 Ti line rendered EVERYTHING .. which means if you were playing a game.. and there was a mountain or a hill in front of you.. or a building.. it would be rendering that building.. plus whats behind it and maybe even behind that. The Kyro chipsets only render what you see.. this gives you astounding video quality and good FPS even in high poly situations.

br0wn is obviously talking about something a lot more complicated .. I'm just opening the curtains to what is already out that does the same task to an affordable degree .

 

[aka1nas]

I think that the Geforce3 and above have Occlusion Culling and will not render portions not currently visible.(Though it may be in software through Det drivers, not sure on that)

 

[Nefrodite]

actually ati also has it to a degree

 

[GoldMember]

I hate ATI like the plague.

 

[br0wn]

Actually, all current modern graphic cards have some forms of occlusion culling.

However, they are not well-suited to handle models with massive amount of polygons yet (at least not in the order that we are dealing).
Why?
Because of the limitation of AGP bus. AGP 4x can handle ~10^9 bytes per sec (this is theoretical number, not even effective bandwidth which should be much lower).

The model we are talking about requires at least 1.8 x 10^11 Bytes per sec (60Hz * 10^9 triangles * 3 bytes/per vertex).

Doing occlusion culling early in the software, and sending just PVS (possible visible set) of the triangles would speed up much of the computation (and allow us to do real-time rendering). In fact, this is what many game-engines have done previously (like quake3). We are just pushing it to the next limit.

Of course, I am not just talking about occlusion culling and LOD, there are many other techniques that we include in the process (they are too technical to be discussed in the open).

 

[GoldMember]

*nods*

I got lost on your math anyway .
I just know simple terms and some in depth but not your math .

 

[dpopiz]

"they are too technical to be discussed in the open" hehehehehehe
so technical they're toxic?