RPU future cards - rennaisance for ray-tracing?

DarkSideA8

Gawd
Joined
Apr 13, 2005
Messages
990
Interesting article in Scientific American shows that future graphics cards may shift from the current technique of rasterizing everything to ray tracing. Ray tracing is not new, but advances in hardware, new algorithms and software make it likely that real-time applications (ie games) could see a dramatic improvement in light and shadows, and overall environmental quality in the near future. The new cards would be called "RPUs" (for some of you, this is old news - but this is 'news' for most of us...)

http://www.sciam.com/article.cfm?chanID=sa006&colID=1&articleID=000637F9-3815-14C0-AFE483414B7F4945

Here's the quote: "The quality of 3-D computer graphics is poised for a quantum jump forward, thanks to speedier ways to simulate the flight of light."
By W. Wayt Gibbs

Ray tracing has been used for a long time, but has generally been limited to high end commercial enterprises (the movie "Cars" apparently took days to process frames, even using Pixar's nifty toys - but they had to use ray-tracing to ensure the reflections off the metal looked realistic). The point of the article is that recent advances may bring this capability down to the level of a high-end PC within the next couple of years.

A more technical look here:

http://graphics.cs.uni-sb.de/Publications/2006/bkd_gh06_final.pdf
 
Good stuff. Thanks for the tip to the other post.

Am I correct in reading that ray-tracing may allow for true round features, not merely simulations using smaller and smaller triangles to hide the fact that what you are really seeing is a bunch of angles?

Is it the rasterization process that requires replacing circles, columns, spheres and cones with multi-faceted polygons?
 
I use POVRAY now and then. Fantastic program, and the price is right.
If vidoe cards can do in hardware what POVRAY does in software at good speed, we'll be really happy.
 
DarkSideA8 said:
Good stuff. Thanks for the tip to the other post.

Am I correct in reading that ray-tracing may allow for true round features, not merely simulations using smaller and smaller triangles to hide the fact that what you are really seeing is a bunch of angles?

Is it the rasterization process that requires replacing circles, columns, spheres and cones with multi-faceted polygons?

Nope, everything is still made up of triangles whether you use ray tracing for lighting or not. As for moving toward ray tracing for realtime apps, I would have to say no again. Say you do todays graphics with a ray tracer, next gen graphics would require an higher order of magnitude of additional processing to improve visuals, and so on for the next generation.
 
DarkSideA8 said:
Good stuff. Thanks for the tip to the other post.

Am I correct in reading that ray-tracing may allow for true round features, not merely simulations using smaller and smaller triangles to hide the fact that what you are really seeing is a bunch of angles?

No, Ray tracing is just about light. In 3D (modeling/animation and games), there is Ray Tracing or Shadow Maps. All a shadow map is is a 2d grayscale texture mapped on to a surface (polygons) and darkens it wherever the texture is darker. All those fancy shadows and lights that are cast all over a game level... those shadows aren't generated real time.. it's faked. The shadows are pre-calculated, unless they are hardware lights. I don't know off the top of my head how hardware lights are handled, but I would be surprised if they are ray traced.

Ray Tracing is different. It takes the light source and generates a number of photons (like say.. 10,000 per light). They are emitted and the 3D engine calculates the path of each photon. It flies out of the light, hits the grey wall (which changes the color and intensity of the photon depending on the wall shader), the photon is bounced off in whatever direction, it hits a glass cup on the table and the photon is refracted as it passes through the glass shader, bounces off the table and etc. until the photon is out of energy, or goes through a certain number of reflects/refracts.

And repeat for every photon, every frame.. Not to mention the processing of whatever animation and physics and gameplay is going on..

It's a very intensive process (which is why it takes so long to render a frame of anything with Ray Tracing in a 3D app) and although graphics processing has come a long way, I think it will be quite a while before we see hardware able to process real-time ray tracing in modern games.

Edit: Darn you Lord of Shadows! You reply faster than I do!!!
 
So, what is keeping truly round or cylindrical objects out of games? Why do we do everything out of triangles?
 
DyerWolf said:
So, what is keeping truly round or cylindrical objects out of games? Why do we do everything out of triangles?

Because we have to. Just as a 2D computer image is composed of pixels, a 3D mesh is composed of triangles. No matter where you position the three points (vertices) of a triangle, it always forms a plane (face) - these faces describe the surface of a given object. The vertices can be assigned colours or to texture/displacement/etc maps in order to make it appear as a solid. More importantly, each plane has a 'normal' which effectively says what direction it faces, allowing light to interact with it and actually make it appear 3D. The only way to make a round object is to increase the number of triangles until it is close enough to being round that the viewer is 'fooled'.
Most importantly, it's how graphics cards and DirectX/OpenGL handle 3D, and this isn't going to change (anytime soon).
 
Lord Twilight said:
No, Ray tracing is just about light. In 3D (modeling/animation and games), there is Ray Tracing or Shadow Maps. All a shadow map is is a 2d grayscale texture mapped on to a surface (polygons) and darkens it wherever the texture is darker. All those fancy shadows and lights that are cast all over a game level... those shadows aren't generated real time.. it's faked. The shadows are pre-calculated, unless they are hardware lights. I don't know off the top of my head how hardware lights are handled, but I would be surprised if they are ray traced.

Ray Tracing is different. It takes the light source and generates a number of photons (like say.. 10,000 per light). They are emitted and the 3D engine calculates the path of each photon. It flies out of the light, hits the grey wall (which changes the color and intensity of the photon depending on the wall shader), the photon is bounced off in whatever direction, it hits a glass cup on the table and the photon is refracted as it passes through the glass shader, bounces off the table and etc. until the photon is out of energy, or goes through a certain number of reflects/refracts.

And repeat for every photon, every frame.. Not to mention the processing of whatever animation and physics and gameplay is going on..

It's a very intensive process (which is why it takes so long to render a frame of anything with Ray Tracing in a 3D app) and although graphics processing has come a long way, I think it will be quite a while before we see hardware able to process real-time ray tracing in modern games.

Edit: Darn you Lord of Shadows! You reply faster than I do!!!


photons... yikes...


pretty sure you more accurately described global illumination type rendering, as opposed to raytracing
 
Lord of Shadows said:
Nope, everything is still made up of triangles whether you use ray tracing for lighting or not. As for moving toward ray tracing for realtime apps, I would have to say no again. Say you do todays graphics with a ray tracer, next gen graphics would require an higher order of magnitude of additional processing to improve visuals, and so on for the next generation.

Um..I think you (and Lord Twilight) have raytracer confused with radiosity, which is usually done as a precomputation step to simulate the ambient lighting in an area. The technique that is normally referred to as raytracing is where you have a screen full of pixels, and then for each pixel you send out a ray where you check to see if/what the ray collides with any object in the environment, and if it does, use data from that intersection/collision to determine what the pixel should be.

DarkSideA8 said:
Good stuff. Thanks for the tip to the other post.

Am I correct in reading that ray-tracing may allow for true round features, not merely simulations using smaller and smaller triangles to hide the fact that what you are really seeing is a bunch of angles?

Is it the rasterization process that requires replacing circles, columns, spheres and cones with multi-faceted polygons?

Ray tracing DOES allow for truly round features, and most demos showcase spheres and such for ray-traced images. However, you can theoretically get this with regular triangle rendering by doing bezier surfaces or something and getting a triangle density that is so high that you get nearly a 1 to 1 correspondence between the geometry onscreen and pixels being rendered to. The rasterization process is where the change would be required for using circles, etc. instead of triangles, however given the extreme simplicity (and hence speed) of rasterizing triangles compared to simple geometric shapes, it's unlikely that we will move away from triangles for typical rendering.

Re: the original topic though, I replied in the previous iteration of this thread and felt that I might as well reinforce the comments I made earlier, as I still believe that they will be true:
Raytracing in games is a long, long way away I think. It's far more likely that GPUs will evolve into extremely parallel, extremely programmable processing units, like a 128-core CPU that will be able to run almost any kind of code that a dev requires: physics, AI, game logic, graphics (either polygonal or raytracing). In the near future though, D3D10, and the chips for it, will be as close to that as we've ever gotten on PCs, and it's likely that some very serious research into doing raytracing on D3D10 GPUs will occur fairly soon. In fact, I think it's very likely that within the next 5 years or so, some developers will be using actual raytracing in games for, at least, either reflections or refractions.
 
Cypher19 said:
Ray tracing DOES allow for truly round features, and most demos showcase spheres and such for ray-traced images. However, you can theoretically get this with regular triangle rendering by doing bezier surfaces or something and getting a triangle density that is so high that you get nearly a 1 to 1 correspondence between the geometry onscreen and pixels being rendered to. The rasterization process is where the change would be required for using circles, etc. instead of triangles, however given the extreme simplicity (and hence speed) of rasterizing triangles compared to simple geometric shapes, it's unlikely that we will move away from triangles for typical rendering.
Exactly! - the stills show glass and mirrored spheres...

I admit I don't know much about the process, but isn't it just math (you know, Pi*r-Squared, etc) to get circular shapes? I've seen the wireframe models used for almost every 3-D application, so I know what the convention is, I just don't understand the reason behind the convention. Is it computing power? Does it simply take so much more computation power to create a curved surface rather than a flat plane?

Or is it that 25 years ago it would have taken more computational power than was readily available and we decided to use an approximation shortcut that has grown into a standard (that might not be relevant anymore)? (for example: in the early days of flight all planes had multiple wings because the mechanics of flight and materials science at the time required the additional surface area to produce lift, at the cost of drag and speed - however "Once thicker wing sections and improved structural materials were introduced, removing the need for external bracing, monoplanes quickly superseded biplanes in aviation. [wiki] - monoplanes being much faster and more efficient than biplanes...)

I mean wouldn't it be simpler to create the wheel on a 3-D model of a jeep by specifying a center point and a radius (for the first circle) and then extrude it, rather than draw an octagon (or higher-faceted polygon) point by point?

Is it the rasterization process thats driving this, or is it all tied together? (triangles~uses~rasterization~reqires~triangles)
 
I admit I don't know much about the process, but isn't it just math (you know, Pi*r-Squared, etc) to get circular shapes? I've seen the wireframe models used for almost every 3-D application, so I know what the convention is, I just don't understand the reason behind the convention. Is it computing power? Does it simply take so much more computation power to create a curved surface rather than a flat plane?

It's a little bit of more than a few things.

One, triangles are FAR more general than spheres and other simple geometric shapes. If I wanted to spend a lot of time (e.g. well over a decade) working on getting something fast, I'd work on something that can be applied to basically 100% of all objects instead of ~1% of all objects.

Two, drawing of triangles is a bit faster to satisfy the level of robustness that is generally required nowadays. Most vertices don't have just position data immediately before rasterization. The data can be stuff like position, texture coordinate, normal, tangent, binormal, or colour, and those have to be interpolated on a per-pixel basis. For triangles it's fairly easy because for the most part, all of the calculations are basically linear (i.e. f(x)=ax+b) but if you were to do that for a sphere, it'd likely get a lot more complex, with a bunch of trigonometry running about, which would run quite slow in comparison. Each piece of data that needs to get interpolated (the example I gave above, which is a bit moderate, would require interpolation of almost 20 floats) has to be interpolated, and a more complex function would reduce the speed at which that can be done.

Three, the way that vertex data is handled now doesn't exactly map well to geometric object data. Right now we store a bunch of vertices and the data for them, and the API automatically translates that data into lists of triangles for very standard use by the hardware.

That's all I can think of now, but I'm sure a more experienced graphics researcher could give oodles more information about the arguments for and against.
 
Back
Top