Ray Tracing: Will It Matter?

[Vinyadan]

I've recently read of this new ray tracing technology, and it looks quite good. However, it also seems to be very, very resource-intensive, more than I think makes sense for mere reflections.

Sooner or later, I'll probably buy a new GPU (there's no hurry, it might take a few months or even be pushed to next year). What I wonder is whether ray tracing could become the new "pixel shading", in that certain (then) new games decided to use pixel shaders that weren't supported on all GPUs, making them unusable.

Nowadays pixel shading is ubiquitous and most games won't work without it, so I wonder if the same will happen to ray tracing in a few years, so that buying a GPU that doesn't support it would be a long-term mistake.

[Mando Knight]

Given that games are made with console capabilities in mind, it makes sense to check in on what the next console hardware will support. Both Microsoft and Sony claim their next consoles will have AMD GPUs with support for raytracing, but until we actually have those consoles we won't know for certain if it's dedicated hardware like nVidia's RTX series or if it's a software approach (such as DXR support for most of nVidia's 10-series GTX cards). Either way, a DXR-capable graphics card will probably be able to handle most games for several years to come even if raytracing support becomes more common in games.

[halfeye]

Raytracing has been around for a long time, you could do it on the Atari ST if you didn't mind waiting overnight to see the results (or so I heard, I didn't try it myself).

Real time is a lot trickier, it might be good, but it'll be a long while until it's cheap enough to be a mass-market thing.

[warty goblin]

I have an RTX card, and it absolutely does make a difference. It's a subtle difference - there's very seldom a single thing that just looks super right with the RTX that was obviously wrong before, but scenes look a lot more naturally dynamic, for wont of a better term. Things look vastly less sterile, and a lot more like they are actually in the space they're occupying. This shows up much better in motion than it does in a screenshot.

I think the more interesting question from a resources point of view is less about the hardware requirements - it's an embarrassingly parallelizable problem and parallelized hardware is still improving pretty rapidly so the necessary hardware will spread pretty rapidly - but a software engineering and game design one. I have no basis for this, but the move from using a lot of very complicated ways to fake the way light works to direct simulation seems like the sort of thing that could actually reduce the difficulty of designing a game in a particular engine. Or it could be an even larger timesink than current methods.

[factotum]

Thinking about regular raytracing, it's not just a "place light sources and objects and everything will magically look correct" thing, unfortunately. For a start, the way raytracing is generally done is the reverse of what you'd expect--e.g. rays are shot out from the "eye" and how they interact with objects and light sources is modelled. This means that transparent objects won't distort incident light as you'd expect, so a lens placed in a scene won't focus light rays. In order to make that work you have to add a pre-processing stage where the software fires light rays from the light sources in the scene at any transparent objects, and remembers what that does to the lighting for the regular ray tracing stage later on--something that I suspect you just can't realistically do in real time.

So, I've no doubt adding this improves realism to some degree, but it's not going to be a "fire and forget" solution, it's going to require a lot of work on behalf of the game artists to get the scenes working at their best.

[Fat Rooster]

Thinking about regular raytracing, it's not just a "place light sources and objects and everything will magically look correct" thing, unfortunately. For a start, the way raytracing is generally done is the reverse of what you'd expect--e.g. rays are shot out from the "eye" and how they interact with objects and light sources is modelled. This means that transparent objects won't distort incident light as you'd expect, so a lens placed in a scene won't focus light rays. In order to make that work you have to add a pre-processing stage where the software fires light rays from the light sources in the scene at any transparent objects, and remembers what that does to the lighting for the regular ray tracing stage later on--something that I suspect you just can't realistically do in real time.

So, I've no doubt adding this improves realism to some degree, but it's not going to be a "fire and forget" solution, it's going to require a lot of work on behalf of the game artists to get the scenes working at their best.

One of the useful fundamental physics properties is the non existence of Maxwells demon, which means that light will be distorted the same way either way through a lens or material. You have to be looking at relativistic speeds before the light going the wrong way gives non physical results. Lenses work just fine. Locating paths from the eye to a light is hard, but accelerations exist.

RTX is certainly a big deal, but not just for ray tracing. The 'support' for ray tracing actually comes from extremely efficient neural network performance. Ray tracing results in very noisy images without huge numbers of rays, so typically ray traced images are passed through a AI denoiser. The RTX cards have very good support for this denoiser, making ray tracing their mass market selling point. Selling your cards as "supports Skynet" might not have the same effect, even though their machine learning support is a leap forward.

[Vinyadan]

RTX is certainly a big deal, but not just for ray tracing. The 'support' for ray tracing actually comes from extremely efficient neural network performance. Ray tracing results in very noisy images without huge numbers of rays, so typically ray traced images are passed through a AI denoiser. The RTX cards have very good support for this denoiser, making ray tracing their mass market selling point. Selling your cards as "supports Skynet" might not have the same effect, even though their machine learning support is a leap forward.

Does that work through the cloud, or do the GPUs contain a neural network?

[factotum]

One of the useful fundamental physics properties is the non existence of Maxwells demon, which means that light will be distorted the same way either way through a lens or material. You have to be looking at relativistic speeds before the light going the wrong way gives non physical results.

That's not the problem. The problem is, in order to get everything to work absolutely perfectly you would need an infinite number of light rays, which, as I'm sure you'll understand, would be computationally intensive. Generally ray tracers use only one ray per onscreen pixel, and so a lot of interactions simply don't get modelled--including the lens situation, or the sort of specular effects you see behind a glass of water when a bright light shines on it.