Investigating ray tracing, the next big thing in gaming graphics | Hacker News Reader

Investigating ray tracing, the next big thing in gaming graphics

extremetech.com

32 pointsevo_926 comments

Oh ray tracing, the nuclear fusion of computer graphics. It's coming! It's amazing! Meanwhile the industry still keeps burning coal.

For applications that deal in perfect lighting, yes it's probably a useful technique that would benefit from hardware support. But where is the real-world benefit for the rest of us? For most consumers, lighting provides about 2% of a game's value and 0% of everything else.

So I sympathize with this point of view, but the difference can be pretty stunning. Although between getting ray tracing or getting something like NURBs as the fundamental unit of geometry composition I'd probably lean toward the latter as a bigger 'win.'

One of the things about ray tracing that was unexpected for me was that you could use a lower pixel resolution and get a higher fidelity 'feel' for the render.

Ugh, not NURBS again, they are so much harder to control for high-detail surfaces you see for modern organics, and provide little to no advantages to hard and most environmental surfaces.

Leave NURBS to vehicle and product CAD, where it's actually very useful

Ok, not NURBS but its annoying that I can describe a sphere with 6 NURBS perfectly and can't with 10,000 triangles.

If you're ray tracing, you might as well use a vec4 for that sphere. :)

Good point, it does make somethings easier.

I for one, believe it would considerably simplify the implementation of graphics engines. To make rasterization look good you have to do a lot of optimizations and go trough a very long pipeline. You don't have this problem with ray-tracing. Plus the same base technique (ray-tracing) can be used for a lot of problems in graphics: visibility checking, lighting, reflections, diffraction, motion blur, ...

I've heard as much too. It'd be interesting to see today's best raster effort put against the same game assets in a ray-traced engine with a real-world game going on.

Anything can look good when you've got 10 minutes and 500 processors. I want to see what happens in 16ms. :)

I mean, as a good point, ray tracing still isn't fast enough for the vast majority of pre-computed rendering for motion pictures. See PRMan and the majority of other modern renderers that all use still use the REYES algorithm. Sure, a unbiased renderer is great when you just need a single still without a huge amount of tweaking, but when you need motion ray tracing -everything- will remain impractically slow for a very long time.

Since you're concerned with frametimes, let's also add in the asset production pipeline and effort. As it's been mentioned, cutting edge rasterizing consists of lengthy pipeline for every single effect imaginable(also sometimes to the graphics assets, mind you) while raytracing does all that almost by default(soft shadows and global illumination are problematic for example, path tracing fixes this but is some two orders of magnitude slower).

The real problem with raytracing is that it's asymptotic computational complexity is far worse when you count scene objects. To compensate this you either lower the resolution or lower the lifetime of a ray(how many bounces it's allowed to make).

Very interesting. Thank you for the additional detail!

Sure, it's awfully simplifies everything, but at a huge cost.

> lighting, reflections, diffraction, motion blur

Just about every one of those becomes easy, but incurs a huge performance penalty.

Also, thinking about it, I'm not sure that motion blur actually becomes significantly easier.

Lighting isn't a huge win, direct lighting is improved slightly, but GI is still impractically slow.

Reflection and diffraction are a pretty good win, but are also things that are solvable with a "ray-trace sometimes" approach, which is a much more likely solution for the near to mid-term.

And generally speaking visibility is done with some sort of ray checking anyways, sure it simplifies things that now we don't have two different solutions, but the complexity is a huge optimization.

So sure, it'd simplify things a ton, but I don't imagine it to be the primary rendering technology in real-time game for a long time.

> Also, thinking about it, I'm not sure that motion blur actually becomes significantly easier.

For motion blur, you take frames at different points in time and distribute the tracing ray over time. So it's actually pretty easy. Don't know if it's practical or efficient tough.

similar sort of thing can be done in a rasterised system, just distribute the rasterised triangles over time, tada.

It would definitely make the engine implementation more simple and elegant, but why exactly would that be worth it? Why would that be worth all the downsides?

I think a good analogy, say that I'm implementing a database. In my database, every single time it handles a query, it always does O(n) search across every record to find matches. The implementation would be so simple and elegant!

Ray-tracing as it currently stands will never dominate computer graphics, because it is essentially brute force. Progress is being made in reconstructing a scene out of a very noisy, low sampling rate [1], which is a much better indication of the future direction of raytracing, where a series of advanced prediction and reconstruction algorithms are used to efficiently render scenes, rather than simply throwing more hardware at the problem.

[1] http://www.tml.tkk.fi/~samuli/publications/lehtinen2012siggr...

I've only looked at the images in the PDF, and not the text, but could a similar technique be applied to image demosaicing?

One of the best real-time raytracers around is Brigade 2: http://www.youtube.com/watch?v=QF2c1uHiYyY http://www.youtube.com/watch?v=Qdw1HvzKt1M

It could really use a 3rd GTX580 to run the reconstruction filter that blackhole linked. But, it's getting better at a steady rate.

Meanwhile, the voxel cone tracing technique mentioned in the article (http://www.youtube.com/watch?v=fAsg_xNzhcQ) is a great hybrid. Rasterizers are great at primary visibility, but are terrible when the algorithm needs non-local information. By supplementing the rasterizer with a fast-but-blurry ray-tracer, you get most of the benefit of both techniques.

Increased use of ray tracing effects in games? Sign me up!

Fully ray-traced game engines? Oh no not again. Damn you Nintendo for starting this nonsense with those faked Project Reality demos...

Graphics engines have more in common with street magic than simulation in that the best ones use just enough physics and mathematics to get the job done and not a bit more. Wake me up when John Carmack says ray-tracing is the next big thing.

That said, ray-tracing has all sorts of unconventional uses outside of strict computer graphics i.e.:

http://www.tomsarazac.com/tom/Bicycles/headlight-tracer.html

John Carmack has said a number of times that it is not the next thing, but it is the thing after next (or the thing after).

The benefits of ray tracing come in precisely because it is more like simulation than street magic. Once you get to sufficient complexity, it is too hard to maintain an illusion. The truth is more consistent than an elaborate lie.

We are not yet close to that level of complexity. We will be eventually. Anyone putting a date on that time before it arrives is almost certainly going to be wrong.

This is exciting. Indeed, it has been "just a few years away" for a few decades now but at some point I'm sure it will show up.

I'm sure I'm not the only one on HN who has written a ray tracer, and once you have it's easy to extend it to something like Metropolis light transport or photon mapping. I remember when my code first worked (without error) how amazed I was that such a small amount of programming can achieve such realistic light simulations!

A website known as OMPF used to be about the only forum online for people interested in ray tracing and global illumination. For some reason, it seems as though it has gone down :(

I would like to see somebody take consumer gear and matrix it to the point where it supports real-time fully photorealistic ray-tracing.

Yes, it would be crazy expensive. Yes, it would be impractical. But it would show that a) it is possible to create without inventing anything new, and b) it costs $X dollars. Leave Moore to figure out the rest.

The industry could use a benchmark to tell us just how far away this is and a model of how to get there without creating a billion-core CPU.

Also, it seems like this would be something that would be amenable to a FPGA system. If so, it might be a lot closer than we realize. Don't know.

Uh, you mean like this ? http://www.youtube.com/watch?v=blfxI1cVOzU

Intel was pushing their Larrabee architecture as the answer for a while.

Actually no. I've looked at a couple of these videos, and all they show is that current technology can be pushed to somewhat approximate realtime ray-tracing. Here's another example of that: http://www.youtube.com/watch?v=zbokPe4_-mY&feature=relat...

So I'll spec it out. HD video, 30fps, editable with in-world physics.

Don't see anything off-the-shelf that's close to that today. My question was if it were possible to hack something together. If so, what did the configuration/cost look like.

Perhaps the answer right now is "no". That's cool. Just wondering.

It is certainly possible to hack something together, the national labs have done this for a while using off the shelf parts to create a 'video wall' that allows them to visualize their simulations. Taking the linked example and applying that technique would suggest you could use somewhere between 12 and 16 'copies' of their setup driving each driving a tile of your HD display.

'in world physics' is well managed by things like the PhysX engine that nVidia sells.

I find it strange that people talk about ray tracing as if it is the solution when in fact it's not even the best of the approximations to the rendering equation that we have. For example caustics can't be adequately represented with ray tracing.

Photon mapping, radiosity, etc can give much closer to photorealistic results.

Reply on news.ycombinator.com