Here are few screenshots of the realtime raytracer we developed at VRcontext (www.vrcontext.com), and that has been presented during the Intel Research Days last week. It's a realtime raytracer designed for massive models. The images are from two different models, one is an oil platform made of a bit more than 100 million triangles, and the other one is a chip manufacturing factory of 200 million triangles (you need few gigabytes of memory to keep these in memory), there is no instantiation or level of details, it's all there each frame. On our local 16 core machine it runs at 20 to 30 fps under the following conditions: 640x480 with 4 times supersampling (x4AA), dynamic shadows from a directional light and on the fly real ambient occlusion (not backed, neither fake SSAO). We observed perfect scalability (x15.9 speed up compared to monocore). The demo machine was 24 cores, so we believe the performance was around 37 fps. Without ambient occlusion you can navigate at full framerate (>70 fps). As explained, it runs only on the CPU(s), the GPU is not used at all (not even for image postprocessing).
This type of models are quite unfriendly regarding geometry and topology (like very long and thin triangles mixed with regular ones, clashes and so on) so the kdtree builder has hard work here.
Images are pure white because we didn't have any material information, we basically had a polysoup to work with.

原帖地址：http://www.gamedev.net/community/forums/topic.asp?topic_id=497681