800TFLOPS Multicore IC for Realtime Ray Tracing(techon.nikkeibp.co.jp)
techon.nikkeibp.co.jp
800TFLOPS Multicore IC for Realtime Ray Tracing
http://techon.nikkeibp.co.jp/article/HONSHI/20090629/172373/
11 comments
Thanks for posting, this is very interesting. I wish they included more details.
The realm of realtime computer graphics has been totally dominated by the SGI+Renderman line of thinking, and realtime raytracing has only made occassional appearances in demos or academic research (and possibly the pipe dreams of Intel executives wishing for a disruptive technology to make up for their weakness in graphics).
It's fascinating that they're willing to build custom chips for this purpose. Are Toyota's needs simply so advanced that they couldn't do anything else? Or will this be another Cell processor type project, a chip launched with high hopes and only one major customer?
The realm of realtime computer graphics has been totally dominated by the SGI+Renderman line of thinking, and realtime raytracing has only made occassional appearances in demos or academic research (and possibly the pipe dreams of Intel executives wishing for a disruptive technology to make up for their weakness in graphics).
It's fascinating that they're willing to build custom chips for this purpose. Are Toyota's needs simply so advanced that they couldn't do anything else? Or will this be another Cell processor type project, a chip launched with high hopes and only one major customer?
Automotive applications are different in that they're inordinately concerned with the appearance of reflections, especially in areas such as the appearance of the headlights. This makes non-raytraced approaches substantially less valuable in the design process. They really want to know what the shiny bits will look like ahead of time.
Raytracing still suffers form the performance implications of incoherent memory accesses patterns for non-primary rays, but when you want physically correct reflections and shadowing to fall out of the rendering architecture naturally, an implementation of ray tracing in hardware starts to look like a pretty spiffy idea.
Are there other fields that require physically correct models for reflections, etc? None spring to mind for me right away.
Raytracing still suffers form the performance implications of incoherent memory accesses patterns for non-primary rays, but when you want physically correct reflections and shadowing to fall out of the rendering architecture naturally, an implementation of ray tracing in hardware starts to look like a pretty spiffy idea.
Are there other fields that require physically correct models for reflections, etc? None spring to mind for me right away.
antenna design, but the spectrum is a little different
Interior decoration ?
(No, not kidding - even muted reflections from walls and curtains may possibly play a large factor.)
(No, not kidding - even muted reflections from walls and curtains may possibly play a large factor.)
Designing lenses and mirrors for things like telescopes and lasers, I'd imagine.
Also, a slight adaptation could be good for radiation and electromagnetic research (e.g. how well a material blocks beta radiation, etc.)
Also, a slight adaptation could be good for radiation and electromagnetic research (e.g. how well a material blocks beta radiation, etc.)
It's amazing that these chips aren't sacrificing precision for performance, operating on Bezier curves.
While we're on the subject of realtime raytracing, does anyone have insight about the polygons vs voxels debate for normal (i.e. desktop apps, games) programmers?
While we're on the subject of realtime raytracing, does anyone have insight about the polygons vs voxels debate for normal (i.e. desktop apps, games) programmers?
The last time I heard anything about this, the bottleneck on the voxel side was problems with surface texturing. That was some time ago though, and I'm not sure what progress has been made on that front.
It seems having followed this over the last several months that many argue that ray tracing is really great, and obviously the way the world works, but scene graphs and filters and such with existing matrix transformations like in DirectX and OpenGL are always going to outperform raytracing because it is just inherently simpler. However I don't do too much development in this area, so I would welcome any other perspective!
I don't think it's a matter of performance, but of realism. If you want realistic calculations, you must go to a ray tracing algorithm, there's no substitute.
If you don't need 100% realism, though, you can probably make things look 99% as good with "fake" effects, with better performance.
If you don't need 100% realism, though, you can probably make things look 99% as good with "fake" effects, with better performance.
Correct. The "fake" effects have limitations and you can break them under certain circumstances, so raytracing is always a good choice when you aren't sure how the materials will be used and under what conditions they will be rendered.
Edit, Ok, http://techon.nikkeibp.co.jp/article/HONSHI/20090629/172373/... has 88TFLOPS equivalent per IC which might be possible if equivalent is loosely defined. And with 9 of those you might hit 900TFLOP's equivalent.