Add Another Core for Faster Graphics

Dzonatas writes "Need a reason for extra cores inside your box? How about faster graphics. Unlike traditional faster GPUs, raytraced graphics scale with extra cores. Brett Thomas writes in his article Parallel Worlds on Bit-Tech, 'But rather than working on that advancement, most of the commercial graphics industry has been intent on pushing raster-based graphics as far as they could go. Research has been slow in raytracing, whereas raster graphic research has continued to be milked for every approximate drop it closely resembles being worth. Of course, it is to be expected that current technology be pushed, and it was a bit of a pipe dream to think that the whole industry should redesign itself over raytracing.' A report by Intel about Ray Tracing shows that a single P4 3.2Ghz is capable of 100 million raysegs, which gives a comfortable 30fps. Intel further states 450 million raysegs is when it gets 'interesting.' Also, quad cores are dated to be available around the turn of the year. Would octacores bring us dual screen or separate right/left real-time raytraced 3D?"

Comments

Gaming

[Anonymous Coward]

There are already ray traced games. :O

http://graphics.cs.uni-sb.de/~morfiel/oasen/

It's been done...

[SigILL]

F.A.N. released a real-time raytraced demo at breakpoint back in 2003. It does no more than 10 fps on my lowly 1GHz P3, but I'm sure it runs quite smooth on a nice modern CPU (though I don't think it's multithreaded).

Put it on the GPU

[TheRaven64]

The thing about ray tracing is that it's the archetypal embarrassingly parallel problem that makes heavy use of floating point arithmetic. The thing about GPUs is that they are incredibly parallel processors optimised for for floating point operations.

Take a look at the proceedings from any graphics conference in the last three or four years, and you will see several papers which involve ray-tracing on a GPU. Actually, not so many recently, because it's been done to death. The most impressive one I saw was at Eurographics in 2004 running non-linear ray tracing. As the rays advanced, their direction was adjusted based on the gravity of objects in the scene. The demo (rendered in realtime) showed a black hole moving in front of a stellar scene and all of the light effects this caused.

Re:Put it on the GPU

[smallfries]

The problem with raytracing researchers is that they are incredibly myopic. *Everybody* should use raytracing for *everything* because it is superior to raster in *every case*. Well, bullshit. Take a look at the raytracing results people have posted links to, and then watch the video of Crysis. The problem is not raytracing, but geometric complexity. Raytracing does not scale nicely with the amount of geometry - mainly because of the shadow rays that have to be scattered from each intersection. The 100mil figure assumes about 100 rays per pixel. Well, you need 64 of them just to get around aliasing, and that doesn't leave many for ambient and shadow bounces.

But the GPU is interesting for raytracing. As it moves closer towards a giant floating point vector machine the motivating application will become raytracing. So at the moment a 7800gtx can push 280Gflops. That is 2800 cycles per ray for a single frame. (BTW Intels figures in the article are bullshit. 100mil rays at 30fps = 3 billion rays per second. Roughly one ray per cycle on averge. They are counting a huge number of rays that have been optimised out of the scene, eg shadows or interpolated from pervious frames using a cache).

The raw horsepower is getting there on the card but at the moment the communication soaks up all of the time. Raytracing is the poster-child problem for parallelisation - assuming that you have random access (readable) global memory. If you need to partition the memory into the compute nodes it begins to get harder. In a GPU building datastructures to hold the information is the bottleneck, and it drops the speed by factors of 100s or 1000s. Nvidia and ATi have given the general-purpose community hints that they will improve performance in reading data-structures so this particular roadblock may disappear. A real scatter operation in the fragment shader would be nice, but you would have to gut the ROPs in order to do it. This may happen anyway as the local-area operations that the ROPs compute could fold into fragment operations. To increase the write bandwidth in the card the retirement logic needs to start retiring 'pages' of pixels anyway, over a much wider bus. Otherwise the number of feasible passes per pixel will always be capped by the speed that the ROPs can retire the data.

So given how hard it would be to *efficiently* raytrace on a GPU - why bother when you can throw so much more raw horsepower at faking it with cheap raster techniques?

Quake 3: Raytraced

[Anonymous Coward]

Just found that game using raytracing - Quake 3: Raytraced.
http://graphics.cs.uni-sb.de/~sidapohl/egoshooter/

Rumors are there's a q4 version on the way.

If you can't beat them, obviate them!

[DoofusOfDeath]

I wonder how much this research relates to Intel's renewed desire to become a graphics player.

If they're having trouble, for staffing or other reasons, producing good GPU designs, then it would be pretty clever of them to revolutionize the industry AND capitalize on their CPU strengths in a single move. More power to them, I say. (More power = about 120 watts, I'm guessing.)

It's not JUST FP that's the issue

[N Monkey]

The thing about ray tracing is that it's the archetypal embarrassingly parallel problem that makes heavy use of floating point arithmetic. The thing about GPUs is that they are incredibly parallel processors optimised for for floating point operations.

It's not just the sheer number of FP calculations that can be the problem. Once you get away from the first (or perhaps even second) level of rays, you end up losing coherence between neighbouring rays which causes memory page/cache thrashing. This is not a nice thing on a GPU.

Re:"entirely vectors"

[S3D]

No, ray tracing is all about searching databases for ray-object intersections. That's what GPUs can't do at all.

Serious raytracers are tile-based anyway, that is using a lot of look-up tables. Processing of single tile could probably be fit into upcoming GPU with "unified shader architecture". But it wouldn't be efficient. GPU arn't designed for a lot of branching.

30 fps - unlikely

[DrXym]

Ray tracing works by tracing a hypothetical ray(s) of light back from a screen pixel, and following it as it bounces and splits off various objects which may or may not be opaque, shiny, textured etc. to the light source. So a ray might first hit a sphere so you calculate the light at that point and recursively to trace the light as it bounces off other objects. To get any level of realism you're talking about multiple recursion which takes an enormous amount of time in any complex scene. Transparency also requires the relected and refracted ray to be traced so the number of rays can increase dramatically.

Ray tracing also suffers terribly from "jaggies". Edges look bad because rays can just miss an object and cause really bad stepping on the edges of objects. To eliminate jaggies and do anti-aliasing, you need to do sub-pixel rendering with jitter (slight randomness) to produce an average value for the pixel. So you might have to trace 4 or more rays in a pixel for acceptable anti-aliasing. Effects like focal length, fog, bump mapping etc. cause things to get even more complex. Most pictures rendered with high quality on Blender, POVRay etc. would take minutes if not hours even on a fast / dual core processor.

The only way you'd get 30fps is if cut your ray trace depth to 1 or 2, used a couple of lights, cut the screen res down and forgot about fixing jaggies. It would look terrible. Oh and find time for all the other things that apps and games must do.

Note that OpenRT is not open source

[Anonymous Coward]

It's nice that people are working on ray traced games, but please note the following:

Oasen is based on "OpenRT" --- which is entirely proprietary, and is NOT open source. Their FAQ explains that clearly.

I'm sure that I'm not the only person annoyed at their use of "open" to mean "closed".

Time to look for an open-source raytracing engine designed for interative use ...

Re:Not quite

[tgd]

Thats because a reflection creates another ray segment, and a refraction creates two.

Considering a non-reflective ray traced world at 800x600 needs 320,000 rays to be cast to calculate an image, so 9,600,000 at 30fps, the claim of 450 million ray segments makes sense... thats 45+ per pixel at 800x600, which is a lot of reflections. Usually you'd limit the number to a fairly low because 100 deep reflections don't add noticable detail, especially in motion. Thats a lot of room for both refractive and reflective objects to be in the scenes.

Re:Not quite

[TheRaven64]

You probably wouldn't just use one ray per pixel. It is typical to fire a number of rays and then average the result. This is because rays diverge quite quickly after passing through the display port, and so you get quite an uneven image. There is a noticeable difference between 1 and 4 rays per pixel, and between 4 and 9. After 9, you start to get into diminishing returns, and beyond about 25 it becomes harder to spot the difference (note that it is common to use a square number of rays, since that makes it easy calculate where they should go).

Re:Not quite

[TheRaven64]

What really needs to be done is to track motion the way you would encoding for mpeg, and focus more ray casts in areas of low motion

There was a paper published a couple of years ago (at Eurographics?) about this. Each ray was independent, and would return a value at each intersection (i.e. you get the primary ray value quickly, and then refine it further with secondary, tertiary, etc ray data). When a ray was no longer lined up with a pixel, it was interrupted and terminated. This meant that you got a fairly low quality image while moving quickly, but a much better one when you let they rays run longer. I found it particularly interesting, since it completely removed the concept of a frame; each pixel was updated independently when a better approximation of its correct value was ready, giving a much better degradation.