The only true path of future gaming; Raytracing

[Obsoleet]

With graphics advancements plateauing for quite some time now, I think it's more and more plausible that raytracing may finally start to get its time in the sun.

I was playing Enemy Territory Quake Wars from 2007 a few weeks ago, and the graphics are more than acceptable even today. It definitely did not feel or look like a game from 7 years ago, and we've had a lot of improvements since then. Same with Team Fortress 2, still acceptable. It's really about gameplay, especially once you reach ~2007 level of graphics advancement. Titanfall uses Source.

I'd place money on raytracing fitting well into AMD's APU / HSA strategy. A hybrid rasterizer / ray-tracing engine utilizing HSA makes a lot of sense in a transitional period, with AMD (and Intel, once they adopt HSA) being in a unique position to leverage such a hybrid in the near-future.

That's why I'm so bullish on AMD these days, I think they are on the right track and frankly few others, other than Intel can even follow them down the HSA path.

 

[tweakboy]

I think todays Haswell 4770k can render ray tracing @ 60fps , 780 1x But problem is its not a easy thing. Also ray tracing wont bring a rich environment like Far Cry 2 or Crysis or BF4 .... that is picture quality,, ray tracing is not.

 

[OCGuy]

Ray tracing with weak APUs? Ehhhh not buying it.

 

[DigDog]

RT is one thing in computing that gives me a headache when i research it. half the people working on it say it's the holy grail, made of bacon and hooker tears. look, just click this button and instant lifelike graphics, all on a pentium 3!
other half claim it's a sack of horse manure that will drain the power out of gotham city just to render a marble and a ray of light.

i've seen videos on YT which are quite amazing - and the claims made about them makes me think there's an italian guy working on them in his spare time (when not doing cold fusion).
ofc the other half the comment are just "faked".

tbh i haven't got the time to learn it for myself, just to decide whether it's real or fake. until i get a RT game i can play on my PC, it's a myth.

 

[monstercameron]

dedicated caustics raytracer cards did exist
http://www.solidsmack.com/cad-desig...stic-series2-ray-tracing-acceleration-boards/

http://www.extremetech.com/extreme/...ally-moves-us-towards-real-time-ray-tracing/2

http://www.cgw.com/Press-Center/Web-Exclusives/2013/Review-Caustic-Series-2-RTUs.aspx

http://www.3dartistonline.com/news/2013/07/review-causticseries2-r2100-vizualizer/


also, I have read somewhere that path tracing engines dont need fast memory, can any one confirm?

 

[Jodell88]

dedicated caustics raytracer cards did exist
http://www.solidsmack.com/cad-design...ration-boards/

http://www.cgw.com/Press-Center/Web-...es-2-RTUs.aspx

http://www.3dartistonline.com/news/2...00-vizualizer/

http://www.extremetech.com/extreme/1...-ray-tracing/2


also, I have read somewhere that path tracing engines dont need fast memory, can any one confirm?

None of your links work for me.

No, they don't need fast memory, they just need to hold the entire scene in RAM and work from there.

 

[Jodell88]

Also, Ambient Occlusion falls under the ray tracing umbrella even though it isn't used in unbiased renderers. Therefore Nvidia has been using raytracing technology even though AO isn't accurate.

 

[monstercameron]

None of your links work for me.

No, they don't need fast memory, they just need to hold the entire scene in RAM and work from there.

try again...

 

[Jodell88]

try again...

:thumbsup:

 

[Jodell88]

dedicated caustics raytracer cards did exist
http://www.solidsmack.com/cad-desig...stic-series2-ray-tracing-acceleration-boards/

http://www.extremetech.com/extreme/...ally-moves-us-towards-real-time-ray-tracing/2

http://www.cgw.com/Press-Center/Web-Exclusives/2013/Review-Caustic-Series-2-RTUs.aspx

http://www.3dartistonline.com/news/2013/07/review-causticseries2-r2100-vizualizer/


also, I have read somewhere that path tracing engines dont need fast memory, can any one confirm?

Very impressive piece of kit. However, from a gaming perspective it's still way too slow.

If prices can come down to couple hundred dollars, I'll pick one up for sure.

 

[monstercameron]

Very impressive piece of kit. However, from a gaming perspective it's still way too slow.

If prices can come down to couple hundred dollars, I'll pick one up for sure.

imagination bought them and integrated this in the new wizard uarch, the $2500 card can only push 100M ray/s allegedly at 30W, meanwhile the IMG has their new mobile ready gpu that can do 300M ray/s...

 

[antef]

It was a valid response. What he is saying is that time is better spend on other activities currently. Graphic cards are not calculating magic, they are calculating numbers. Even with hardware acceleration specific to this task, does not mean that the total amount of numbers to be crunched changes. Currently, GPUs are limited to a finite set of limiting resources. Size, heat, material ect.

Great lighting would be nice, but other effects are more important for most people. Once the cost of ray tracing is reduced in relation to other factors, you will see it in games. Right now, it would be too much at the current limits of GPUs. Maybe the new node will change that, but I doubt it.

You still are not reading what I said. Jodell88 is telling me why it's not used in games (because the tech is too slow) and you're telling me why they don't dedicate hardware to it on GPUs for games (because that die area needs to be spent on other things). I wasn't talking about games in my post.

My post pointed out the number of applications where ray tracing is already used. Its utility or feasibility in these contexts is not in question. I'm asking, for these applications, why don't the big GPU vendors produce a special card (not stealing away die area/memory from an existing GPU) that has silicon dedicated to accelerating ray-tracing. It's already been shown by cards like the CausticOne that this does indeed speed it up a good deal, and Imagination is now saying the same thing. Since this is in the GPU space, you'd think the GPU vendors would have hardware that does this well and sold to these customers. Instead, these customers are running their ray-tracing routines on normal CPUs. If Imagination thinks dedicated silicon can do it so well that you will approach near-real-time processing, then I imagine a big dedicated card that did only that would sure speed it up for server farms or 3D/CAD workstations doing this all the time.

 

[BrightCandle]

Right now you can get better lighting effect from using Shaders than you can ray tracing. So the die area is better spent enhancing programmable shaders on the GPU instead of dedicated fix function ray tracing. However at some point in the not too distant future developers will be wanting to increase the accuracy of the lighting and at that point it might be faster to use ray tracing for the same effects. But we'll probably see high quality global illumination long before ray tracing goes mainstream.

 

[Jodell88]

But we'll probably see high quality global illumination long before ray tracing goes mainstream.

Ray tracing is the only way I know of that you can get high quality global illumination. If there are other ways, please share.

 

[Obsoleet]

Ray tracing with weak APUs? Ehhhh not buying it.

Because they'll never advance at all. They're already making APUs that run 720 or 1080P res in new games with no AA/AF. You may not be able to see it now, but Intel/AMD 'APUs' are going to change everything.

 

[TreVader]

I think the improvements to games graphically have and will continue to slow over time, the place where games can really improve is scale. These days the vast majority of games (I'm talking mostly multiplayer FPS) are very limited in scale and really haven't improved all that much over games in the past.


I want to see online multiplayer FPS games that have 50v50 or 100v100 servers and run smooth. Graphics are awesome but I wonder at what point does stuff just look GOOD ENOUGH?

 

[f1sherman]

I think the improvements to games graphically have and will continue to slow over time, the place where games can really improve is scale. These days the vast majority of games (I'm talking mostly multiplayer FPS) are very limited in scale and really haven't improved all that much over games in the past.


I want to see online multiplayer FPS games that have 50v50 or 100v100 servers and run smooth.

You mean like in Planetside 2?
LOL

But how about 7,000+... no, not batches - players(!)

https://www.youtube.com/watch?v=RCK-E5AopVI

B-R5RB BY THE NUMBERS

· Around 21 hours of total fighting
· 7,548 unique characters belonging to those two coalitions participated in the overall battle (i.e. landed on at least one killmail). 6,058 participated directly in the B-R5RB system with 2,670 in system at max
· 717 unique player Corporations
· 55 unique player Alliances

Totals destroyed:

· Titans – 75 (74 in system, one on its way to the fight) N3/PL lost 59 titans and CFC/DTF lost 16 titans
· Titan losses by type: Gallente Erebus – 37, Amarr Avatar – 25, Minmatar Ragnarok – 13, Caldari Leviathan - 0
· Supercarriers – 13 (12 in system, one as it tried to escape the system)
· Dreadnaughts – 370 (356 in system, 14 in connected skirmishes as both sides attempted to stop the other from bringing reinforcements)
· Carriers – 123 (109 in system, 14 in connected skirmishes as both sides attempted to stop the other from bringing reinforcements)
· And lots more smaller ships and probably a bajilion drones and fighters

In comparison, the previous record for largest single battle Titan losses was a tie between battles in “O2O” and “Uemon”, with 12 total Titans destroyed in each.
Approximately 775 doomsdays were fired, which is about 24% of all the doomsdays fired in the last two years inclusive.
The Battle for HED-GP, which preceeded this one in the Halloween War, had about 200 doomsdays.

DISCLAIMER: Apparently, it doesn't run all that smooth. Mostly due to code from 2001.

 

[BFG10K]

We see this almost every year where ray tracing is going to "revolutionize" the market, but nothing ever materializes. How many billions did Intel spend on trying to push it through Larrabee? All we ever got was some Quake 3 implementation, and Quake 4 performance that was too slow for an already old game.

It's similar to that "infinite geometry voxel demo" that keeps popping up every year.

Pure ray tracing has far too many unsolved performance issues to ever be performance-competitive with rasterization.

 

[Mand]

As an optical engineer who uses ray tracing software professionally for design, I can say that it is without a doubt a computationally expensive process. Particularly since you're not just tracing one ray to see where it goes, ideally you want to replicate the diffuse reflectance and scattering properties of things. But that means multiplying the rays, as you trace them. You get away from a ray trace and get closer to a stray light calculation, which is massively intensive due to the sheer quantity of rays involved. It's an exponential progression on each surface, for each ray you trace (e.g. one ray hits surface, spawns ten rays in different directions off of the surface, each of those spawns ten rays when they hit something else, etc). There's just no way to do that with fidelity and keep the processing load down. There are some ways to make it not blow up your computer, but they only reduce the problem, not eliminate it.

It's expensive, but it's also really important. If you want to model lighting properly, you have to start with modeling light properly. The performance they show is really impressive, and I'd love it if we had enough chops in our hardware to pull it off realtime in games.

 

[BrightCandle]

A modern PC game might have 30-40 lights on screen at least, if not many more than that if individual particles are lighting the scene in an explosion. With ray tracing every pixel needs to determine its visibility to every light source as a first order pass, which results in pixels on screen * rays. Then you add into that the fact that direct rays are no where near enough, because light bounces around a scene so you need to do indirect rays as well, and lets assume we only do one such calculation per pixel there per light we get twice as many rays. Practically you need hundreds of rays per pixel on the screen to get any reasonable sense ambient lighting and probably thousands to get a genuine sense of soft shadows.

What does this come out as on a desktop per frame?
1080p = 2073600 pixels
Just 1 direct light = 2 million rays per frame, at 60 fps = 124 million rays cast
40 lights, 1 ray per light = 40 * 2million = 82 million rays per frame at 60 fps = 5.0 billion rays
40 lights, 100 rays for ambient = 40 * 100 * 2m = 8.2billion, at 60 fps = 497billion rays

The hardware they have produce does 30 million rays? That is good but its not enough to make a convincing lighting scene, and even I have underestimated what is truly required for lifelike lighting, 100 rays is no where near enough for really good convincing ambient lighting and in this case I have assumed indirect rays are as cheap direct ones but that isn't the case. Ideally you want light to do more than 1 indirect jump as well, it needs to be based on the strength of the light source how many jumps it makes before there isn't any point adding it. Suffice to say its a computationally heavy process.

 

[realibrad]

You still are not reading what I said. Jodell88 is telling me why it's not used in games (because the tech is too slow) and you're telling me why they don't dedicate hardware to it on GPUs for games (because that die area needs to be spent on other things). I wasn't talking about games in my post.

My post pointed out the number of applications where ray tracing is already used. Its utility or feasibility in these contexts is not in question. I'm asking, for these applications, why don't the big GPU vendors produce a special card (not stealing away die area/memory from an existing GPU) that has silicon dedicated to accelerating ray-tracing. It's already been shown by cards like the CausticOne that this does indeed speed it up a good deal, and Imagination is now saying the same thing. Since this is in the GPU space, you'd think the GPU vendors would have hardware that does this well and sold to these customers. Instead, these customers are running their ray-tracing routines on normal CPUs. If Imagination thinks dedicated silicon can do it so well that you will approach near-real-time processing, then I imagine a big dedicated card that did only that would sure speed it up for server farms or 3D/CAD workstations doing this all the time.

So then you are just asking why dont we see hardware manufactures building hardware for a very very limited market? I suppose the answer would be that its due to the market being very very small. There are applications that use ray tracing sure, but hardware manufactures get more out of investing in other areas than R&D on ray tracing hardware.

 

[beginner99]

A modern PC game might have 30-40 lights on screen at least, if not many more than that if individual particles are lighting the scene in an explosion. With ray tracing every pixel needs to determine its visibility to every light source as a first order pass, which results in pixels on screen * rays. Then you add into that the fact that direct rays are no where near enough, because light bounces around a scene so you need to do indirect rays as well, and lets assume we only do one such calculation per pixel there per light we get twice as many rays. Practically you need hundreds of rays per pixel on the screen to get any reasonable sense ambient lighting and probably thousands to get a genuine sense of soft shadows.

What does this come out as on a desktop per frame?
1080p = 2073600 pixels
Just 1 direct light = 2 million rays per frame, at 60 fps = 124 million rays cast
40 lights, 1 ray per light = 40 * 2million = 82 million rays per frame at 60 fps = 5.0 billion rays
40 lights, 100 rays for ambient = 40 * 100 * 2m = 8.2billion, at 60 fps = 497billion rays

The hardware they have produce does 30 million rays? That is good but its not enough to make a convincing lighting scene, and even I have underestimated what is truly required for lifelike lighting, 100 rays is no where near enough for really good convincing ambient lighting and in this case I have assumed indirect rays are as cheap direct ones but that isn't the case. Ideally you want light to do more than 1 indirect jump as well, it needs to be based on the strength of the light source how many jumps it makes before there isn't any point adding it. Suffice to say its a computationally heavy process.

true but I understood the new PowerVR chip will can be used to only use raytracing for some parts of the scene, metallic surfaces and not for the whole scene. It was 300Mio rays per second but at 60 fps yes, thats not very much at all per image.

 

[PrincessFrosty]

Ray tracing and approximations of it are nice and all, but they're so computationally expensive that traditional rasterization can mimic the ray tracing effects with a high degree of accuracy and still be faster rendering.

Most of graphics up until now have been about getting fast approximations of effects and simply getting more clever about that, we're still at the point where we can make more significant visual enhancements with rasterizers for the same power it would take to do ray tracing in even basic scenes.

You can already do low resolution ray tracing on fast GPUs you just have to have very little scene detail, and that's the problem once you consider the trade off it simply looks worse.

Maybe when we've gone through another factor of x50-x100 increase in power and the retruns on rasterization have become negligable, we'll see a case for raytraced games, we're still a long way off.

 

[BrightCandle]

Most of the lighting we see in todays games is prebaked ray tracing. The lighting that isn't is incredibly simple algorithms like ambient occlusion, bloom, HDR and approximations of fire lighting with close objects. Ambient lighting is baked into the world and then they layer the lighting effects over the top. I guess this is also where they are going with this ray tracing, its something you layer over the top of a prebaked ambient lit scene.

 

[Sable]

We see this almost every year where ray tracing is going to "revolutionize" the market, but nothing ever materializes. How many billions did Intel spend on trying to push it through Larrabee? All we ever got was some Quake 3 implementation, and Quake 4 performance that was too slow for an already old game.

It's similar to that "infinite geometry voxel demo" that keeps popping up every year.

Pure ray tracing has far too many unsolved performance issues to ever be performance-competitive with rasterization.

A thousand times this. ^^^^