peterb@pbear.UUCP (03/01/85)
IF you plan to use hypercube for image processing (i.e. computer graphics for a movie like Star Wars) then why not scrap the cube in any of its designs and go to a 1 to 1 mapping of pixels per processor. If your talking huge then this is not so impossible. each processor execute a ray-tracing algorithm back to a light source or other known end point and produce the color/brightness for that pixel. This would require one high speed one-way bus from a master object definer to each processor and this is done by having all of the processors listen to the bus at the same time. This is possible since each processor needs the exact same copy of the definition for it to apply ray tracing. Then each pixel can be computed and returned to the imager over a slower speed bus (such as ethernet) where it is converted into a rasterized image. Then the imager sends out the definition of the next frame and the process repeats itself. You could even group processors together so their memory requirements are not gargantuan. If you are talking mega-sized systems, then this configuration does have its merits. For a standard "high-res" CAD picture of 640X400 pixels, this requires 256000 processors, each with about 256kb = 65.536Gb of memory. If you shrink this by grouping together 64 processors per memory array(this assums keeping 64K local) memory requirements go down to 1.024 Gb of memory. Surely this is better than spending megabucks on controllers to talk point to point or groupings. I welcome responses and suggestions and even flames. Peter Barada ima!pbear!peterb
ron@brl-tgr.ARPA (Ron Natalie <ron>) (03/02/85)
> > IF you plan to use hypercube for image processing (i.e. computer > graphics for a movie like Star Wars) then why not scrap the cube in any of > its designs and go to a 1 to 1 mapping of pixels per processor. If your > talking huge then this is not so impossible. each processor execute a > ray-tracing algorithm back to a light source or other known end point and > produce the color/brightness for that pixel. > Henry Fuchs of North Carolina and company has done this. What they've managed to do is construct a memory such that each pixel was capable of processing a symple quadratic equation plus a few relatioal operations at the memory access speed. All kinds of neat applicatios ensued. -Ron