Panel: Rays vs. Rasters

When Will Ray-Tracing Replace Rasterization?

by Jan Hardenbergh

Organizer/Moderator : Brad Grantham

Panelists : Kurt Akeley, Brad Grantham, David Kirk, Tim Purcell, Larry Seiler, Philipp Slusallek

Brad Grantham started of the discussion with the guideline that interactivity was the focus. When things are easy & cheep and "good enough" they will take over. He also observed that in raytracing, the image pixels are the outer loop and the geometry is the inner loop. And the rasterizers were rich in features such as bump mapping, reflections, motion blur, etc.

Philipp Slusallek took the strongest position that raytracing would become dominant because it was the right conceptual model. It supports the basic physics and the algorithm is so simple.

He showed several images and videos. One had all of the usual things you see in such a picture, the mirror ball, reflections, translucency, but also a volume (UNC skull), and the dragon light field. The rays just keep passing through material - even if the material is processed by different renderers. They just work via plug & play. The demos Phillip showed were running at 2 frames per second on fairly vanilla hardware.

OpenRT is an API for ray-tracing similar to OpenGL(tm)

Tim Purcell showed how graphics chips, aka GPUs or VGUs, have become stream processors. The streams flow in and through kernels that have access to global data. This structure can be used to write a ray-tracer on the graphics chip. Tim was the first to say that renderers will use hybrids of rasters and rays. He showed a demo of using rasters for the basic geometry and rays for shadows.

David Kirk started by pondering "People who believe they cannot use computers to do rendering, what do they use, pencils?" He believes that the renderers are hybrids. He believes that GPUs are increasing in power at greater than Moore's law. The is mostly due to the memory coherency, streaming, embarrassing parallelism, and deep pipelining.

Kurt Akeley said that ray-tracing would not overtake rasterizers in his lifetime. looked at other claims: Volume Rendering will overtake geometry (Jim Kajiya, 1991), and IBR will replace 3D rendering. What happens it that the raster/texel engine adapts to absorb those techniques. Half of the NV30 is occupied with the depth buffer and the texture sampling to achieve that.

Rasterizers have become very precise and now use infinitely thin rays to generate samples. The removes the argument that rays are better quality.

Larry Seiler pointer out that Cobol is still around, and more relevant - 2D bit BLTs are what these chips do most of the time and that is not going to change. Neither will geometry and rasters. However ray-trace like techniques are being used to achieve more than straight rasters: NPR, blur, reflection, refraction.

Larry emphasized what others have said: GPU will not become general purpose because we need to keep the streaming model to keep them fast.

Realism is not the end goal - portraying information or telling a story is the goal.

Larry ended with a note from Mark Twain: "Rumors of rasterizations death have been greatly exaggerated!"

During questions, Tom Duff noted that ILM is changing their renderer from having geometry as the inner loop to the outer loop. The must do this because the amount of geometry they need keeps expanding faster than the memory capacity of the hardware.