Point Clouds as Basis for 3D Computer Graphics Breakthrough

[Thel]

Have you seen this?

http://lidarnews.com/point-clouds-as-ba ... eakthrough

[Live-Dimension]

I saw this a while back. It's fake.

Hows the taste of your Red Herring?

Urlz

As for claims that Unlimited Detail can do real-time graphics rendering on a machine with a single-core processor and no graphics card, Nvidia people say they’re sceptical. Searching through trillions of points of data would require large amounts of RAM (random access memory), and Dell isn’t sharing any details on how his algorithm deals with that problem.

Really, "Dell" who has a background in super markets talks a lot about the potential of his tech, but doesn't go into it at all. Not one scrap of information has surfaced anywhere about his 'algorithm'. There is lots of problems he sidesteps and he hasn't provided any evidence to backup his claims. Also notice how everything looks the same? Instanced data can be duplicated a lot faster then trying to recalculate it all.

I could go on and on, but there's plenty of this on the net already. Search for it if you must.

[eatcomics]

God, he keeps saying in the videos, well I watched two, that Oh polygons are bad, you should use point cloud data... Well then, give us some more info on d pyour stuff... And the shadows are messed up in all those videos, so obviously shadow mapping will be more difficult... All the shadows look pixely, vs smooth like in real life... But I guess that's a small price to pay for unlimited pixels... But really this magical sorting algorithm doesn't have much to sort if you can't fit your unlimited points in memory... So how much space do these points take up in memory, apparently they don't take up any?

[qpHalcy0n]

Whether this guy has actually managed to do this is something I'm still a little skeptical about. He's a little passive aggressively arrogant in a way only a brit can be ;] I kid.

However arrogant he is, though, this technology is not far off the mark. There have already been several proof of concept demos out that deal with doing this sort of thing in real time. The techniques themselves have been around for decades but we're just now approaching a point where the hardware SHOULD be able to do this. It's not likely that "point cloud" data will win out though.There's a hell of alot of obstacles in the way there. There is hardware in development now that targets ray casted/voxel rendering as opposed to a traditional full fledged rasterizer (see Larabee). Again, this isn't likely to win out though because for the most part there's only a tiny handful (fewer than 10) developers who will actually be able to deal with that radical of a departure from tradition. For those reasons, nVidia and ATi won't hop on the wagon.

The real deal here though is that the future isn't likely to take this form or even a ray-caster that samples from texel sized polygons but would probably be a blend between tessellated and voxel geometry where optimally the geometry would have a resolution of around a texel. So essentially you'd be sending polygons of sizes a few texels or smaller down the pipeline and you'll be able to reap all the benefits of current hardware while simultaneously seeing the benefits of a "theoretically" infinite geometric resolution. In reality, the resolution is only as good as your display which so far as your eyes are concerned is almost infinite.

The memory requirement here is really a non-issue because with these partitioning techniques, similar to texture partitioning techniques, the data exists on disk. Not on chip. It's not something where you just load up these massive textures at run time and clobber your memory with gigs of data, which obviously would be a huge bandwidth nightmare. This would be the traditional way of doing it. Off balance data partitioning and good job scheduling can make small beans out of massive chunks of data so that the "jobs" churn away at data while all the other little threads run out of sync with that. So there's a certain amount of prediction there which is okay because as it turns out there's usually more than enough memory to deal with the unexpected. In the case of virtual texturing (which is down a similar vein), it also turns out to be more efficient because these techniques make use of more efficient and less aggressive sampling routines which take a load off the pixel processors while attaining "theoretically" infinite texture resolution.

Nut in a shell: Don't be surprised to see things similar to this popping up here and there in the near future.

[eatcomics]

Cool qp that was very informative :D