spacemonkey
Newcomer
Nothing in the world moves - just the camera 
Voxel rendering (formerly Death of the GPU as we Know It?)
- Thread starter sqrt[-1]
- Start date
Yeah, I noticed the same thing. No grass movement, no branches swaying in the breeze, no creatures wandering around, and what's with the water? If unlimited atoms are possible, why aren't we seeing 'atomic' water?
I know the obvious answer to that, but you'd hope they would temper it a bit in the face of what they're showing.
I know the obvious answer to that, but you'd hope they would temper it a bit in the face of what they're showing.
You tooka da words right'a outta mah mouth.
Also with Tessellation finally gaining some momentum, wouldnt this help out the polygon camp? (then theres the whole Ray Tracing thingy)
The biggest problem in point/voxel based modeling is the huge memory/storage requirement to store the points/voxels. I don't see any proof in these presentations that this problem is solved, quite contrary. All the scenes repeat just a few objects.
The geometry quality is excellent in these videos. With professional artists and better lighting and image post processing systems they could easily exceed the current game graphics quality in small static environments.
However since they do not mention anything about their "revolutionary" data compression algorithm, I am quite sceptical how this technology could be used in a real game. A single high quality point cloud rock can be over 100 megabytes of data. Polygons combined with displacement maps (tessellation for low frequency detail, and pixel displacement techniques for high frequency detail) offer very high compression ratio compared to high polygon meshes (and point/voxel data). Displacement maps are also very easy to stream in and out of memory (virtual texturing or any other existing texture streaming technology can be used).
If their next presentation is about their revolutionary data compression and streaming solution I will be really really interested in this technology. Personally I do not care so much about the inability of animation/skinning with this technology. The technology could be really useful even if used just for the static environment. The system could output it's data to g-buffers (depth + material channels) and dynamic stuff could be drawn on top of it as polygon models. Lighting could be done with deferred rendering (fully dynamic of course). Deferred rendering doesn't care how the g-buffer data is generated. You can mix and match different techniques without any problems.
No matter how things do proceed, I don't think this technology is the death of the GPU. GPU would be still highly useful for deferred lighting, post processing and many other graphics related parallel processing tasks. And the geometry "search" they describe is a highly parallel task as well. 1080p screen has over 2 million pixels to search. Current generation programmable GPUs are really fast in various parallel searching and data mining operations. If you want to do 2 million parallel searches to a completely static structure, optimizing the algorithm for GPU would likely offer a significant boost in performance.
The geometry quality is excellent in these videos. With professional artists and better lighting and image post processing systems they could easily exceed the current game graphics quality in small static environments.
However since they do not mention anything about their "revolutionary" data compression algorithm, I am quite sceptical how this technology could be used in a real game. A single high quality point cloud rock can be over 100 megabytes of data. Polygons combined with displacement maps (tessellation for low frequency detail, and pixel displacement techniques for high frequency detail) offer very high compression ratio compared to high polygon meshes (and point/voxel data). Displacement maps are also very easy to stream in and out of memory (virtual texturing or any other existing texture streaming technology can be used).
If their next presentation is about their revolutionary data compression and streaming solution I will be really really interested in this technology. Personally I do not care so much about the inability of animation/skinning with this technology. The technology could be really useful even if used just for the static environment. The system could output it's data to g-buffers (depth + material channels) and dynamic stuff could be drawn on top of it as polygon models. Lighting could be done with deferred rendering (fully dynamic of course). Deferred rendering doesn't care how the g-buffer data is generated. You can mix and match different techniques without any problems.
No matter how things do proceed, I don't think this technology is the death of the GPU. GPU would be still highly useful for deferred lighting, post processing and many other graphics related parallel processing tasks. And the geometry "search" they describe is a highly parallel task as well. 1080p screen has over 2 million pixels to search. Current generation programmable GPUs are really fast in various parallel searching and data mining operations. If you want to do 2 million parallel searches to a completely static structure, optimizing the algorithm for GPU would likely offer a significant boost in performance.
spacemonkey
Newcomer
I did a little bit of searching and found this...
http://bautembach.de/wordpress/?page_id=7
Dennis Bautembach, from University of Hamburg, shows a neat way of animating voxels in SVOs. In his thesis (PDF) he says...
[1] citing the Atomontage engine.
So it seems like animation is doable, and may not require a gazillion bytes of memory.
http://bautembach.de/wordpress/?page_id=7
Dennis Bautembach, from University of Hamburg, shows a neat way of animating voxels in SVOs. In his thesis (PDF) he says...
[1] citing the Atomontage engine.
So it seems like animation is doable, and may not require a gazillion bytes of memory.
It is so obvious that the algorithm he is using is extremely memory bandwith limited.
His excuse for the repetitive instances of models is just laughable and makes it obvious that his claims are fraudulent all the way. There are thausands of free to use models available!
Why doesn't he show a vast and unique terrain landscape that is very easy to produce? Because this graphics engine is unable to do this at interactive frame rates.
All he is able to achieve is to impress people who have no clue about graphics technology and don't know that there is nothing revolutionary about what he is showing.
Quote from the Video:
"I am Bruce Robert Dell. In 2010 our technology appeared in most of the worlds media. We have found a way to give computer graphics unlimited power."
What a disgusting scam!
His excuse for the repetitive instances of models is just laughable and makes it obvious that his claims are fraudulent all the way. There are thausands of free to use models available!
Why doesn't he show a vast and unique terrain landscape that is very easy to produce? Because this graphics engine is unable to do this at interactive frame rates.
All he is able to achieve is to impress people who have no clue about graphics technology and don't know that there is nothing revolutionary about what he is showing.
Quote from the Video:
"I am Bruce Robert Dell. In 2010 our technology appeared in most of the worlds media. We have found a way to give computer graphics unlimited power."
What a disgusting scam!
Minecraft developer: It's a scam!
And also, But Notch, it's not a scam!
Assuming their technology is based on sparse voxel octrees, or is reasonable similar to SVOs...
However the data in hard drive(s) would get huge if all that detail was unique. Rage requires 3 DVDs and it only has unique (pretty low res) texturing everywhere. A full game world with unique geometry detail everywhere would exceed all storage media sizes by far. Maybe we could host it in a huge data center and all players could stream it. The bandwidth required isn't that high. It could be feasible in the near future.
SVOs can basically provide "unlimited detail", since you can link child nodes to their parents causing unlimited detail recursion. However this is only good for repeating patterns, not for unique detail.
SVO has automatic level of detail. The further away the ray traverses, the lower quality data is used (we do not need to consider details that are smaller than one pixel). Just like in virtual texturing (or any mipmapping based graphics hardware), the required detail level (and required bandwidth) drops very rapidly as you start processing data further away from the camera. The working set in main memory can be surprisingly small. Similarly virtual texturing requires only 4096x4096 pixel cache to hold the current working set (for 720p resolution). Virtual textured game can manage with just 50 megabytes of graphics memory for all the textures. Similarly SVO can keep very narrow data set in memory, and stream new data on demand.
However the data in hard drive(s) would get huge if all that detail was unique. Rage requires 3 DVDs and it only has unique (pretty low res) texturing everywhere. A full game world with unique geometry detail everywhere would exceed all storage media sizes by far. Maybe we could host it in a huge data center and all players could stream it. The bandwidth required isn't that high. It could be feasible in the near future.
I agree with most of his points. However SVOs only store voxels at the surface, not inside the objects. This cuts down his estimated memory requirement drastically. Also storing voxel colors as 24 bit uncompressed data sounds like overkill to me.
SVOs can basically provide "unlimited detail", since you can link child nodes to their parents causing unlimited detail recursion. However this is only good for repeating patterns, not for unique detail.
Last edited by a moderator:
Which is probably the main reason why Carmack has abandoned the tech after testing it in 2008-2009. Most likely all their trouble with the Rage datasets was another factor.
Virtual texturing and tessellation/displacement would only require another texture layer, 16 bits are more than enough; I'd consider it a far more efficient representation of geometry data as well.
Especially because you can deform this geometry easily, and just because you can bake lighting into the color layer you don't necessarily have to, it can also work with dynamic lighting too.
Then again maybe it is really better for some stuff, we'll see once we have at least 100 GB of background storage for a game...
A good white paper about CUDA accelerated SVO rendering:
http://www.tml.tkk.fi/~samuli/publications/laine2010i3d_paper.pdf
They use DXT-like compression to compress the color data, and have a memory efficient layout for the octree (just a few bits per node). The data sets are several gigabytes each. A full game world would need more sophisticated compression to fit the whole game world in a reasonable space. Their GPU SVO raycaster is faster than their polygon based GPU raycaster, so the performance is pretty good. Additional compression technology could make this technology viable in some next generation console games (not all games require huge game world with lots of unique detail).
http://www.tml.tkk.fi/~samuli/publications/laine2010i3d_paper.pdf
They use DXT-like compression to compress the color data, and have a memory efficient layout for the octree (just a few bits per node). The data sets are several gigabytes each. A full game world would need more sophisticated compression to fit the whole game world in a reasonable space. Their GPU SVO raycaster is faster than their polygon based GPU raycaster, so the performance is pretty good. Additional compression technology could make this technology viable in some next generation console games (not all games require huge game world with lots of unique detail).
The only thing that is going to kill the GPU as we know it is when PC gaming is no longer viable and AMD/NVIDIA are forced into developing console-only parts.
sebbbi, thanks. However I still think rendering is not even the worst hurdle to overcome. If a game has to have polygon models for animating/dynamic objects it kind of defeats the purpose. You have an incredibly detailed terrain with individual pebbles and stuff and right on top you have a 10k poly automobile.
I'm still too fresh into voxels so anyone feel free to correct me, but my intuition is that better storage/retrieval algorithms are subsumed by the inexorable advancement of cache/ram/HDD sizes plus bandwidth. New algorithms should concentrate on the dynamism problem.
sebbbi, thanks. However I still think rendering is not even the worst hurdle to overcome. If a game has to have polygon models for animating/dynamic objects it kind of defeats the purpose. You have an incredibly detailed terrain with individual pebbles and stuff and right on top you have a 10k poly automobile.
I'm still too fresh into voxels so anyone feel free to correct me, but my intuition is that better storage/retrieval algorithms are subsumed by the inexorable advancement of cache/ram/HDD sizes plus bandwidth. New algorithms should concentrate on the dynamism problem.
Camrack may not have made such a straight statement about voxels, but the facts are that Rage doesn't use them, Doom4 won't be using them, and whatever their next engine iteration is going to be, it won't be using them either (as it's going to be just an iteration of tech 5 instead of starting from scratch).
Yeah... Listening to his QuakeCon 2011 Keynote, it sounds like he doesn't want to jump into another scenario where he is trying to figure out how to realistically implement the tech. That entire Keynote was basically an extended discussion on the nightmares of making MegaTextures work for a REAL game, not just a theoretical one where ideal conditions are satisfied. I imagine he sort of sat back, looked at the idea of using SVOs and then said "
No."Modern GPUs can do Voxels moderately well honestly... Not incredible, but decently.
As for why they aren't used, there is honestly an entire laundry list of reasons that are readily obvious(Animation, Shading, Data Size, Memory, Streaming Limitation, etc.), and that is before anyone even tries to build a real game using it only to discover, much like Carmack & Company discovered with MegaTexture tech(which is inherently simpler than SVOs would be), that when you hand the tech to some talented designers and artists that they will literally take all your neat and tidy solutions to problems, and then destroy them in a heartbeat without even breaking a sweat. You may think you have it all figured out, but when you go to make a game, they are just going to prove that you didn't even consider 1/1000th of the possible problems that could arise...
Carmack basically said outright that he wasn't even sure that they would be able to pull MegaTexturing off at times... They very nearly did not hit their targets with framerate and quality...
And SVOs would almost certainly be even more of a nightmare.