- 1,061
- United Kingdom
Polyphony Digital Reveals GT Sport's "Iris" Ray Tracing System at CEDEC 2018
- Thread starter GTPNewsWire
- 88 comments
- 15,488 views
- 1,748
- vspectra06
- vspectra06
I'm hoping they're doing R&D on a real-time GI lighting solution to bring back 24h dynamic time-of-day transitions for GT PS5 though.
30fps games are already using light probes and spherical harmonics with the pre-computed irradiance transfers to have dynamic lighting changes with good global illumination light data. Next-gen hardware should be fast enough to do it at 60fps.
30fps games are already using light probes and spherical harmonics with the pre-computed irradiance transfers to have dynamic lighting changes with good global illumination light data. Next-gen hardware should be fast enough to do it at 60fps.
- 1,239
- Italy
PD you are wizards 
- 161
I personally think that the next gen will be more about new experiences through deeper simulation. The PS5 will most likely have a Navi based GPU with capabilities similar to a GTX 1080 (1080ti at best), so raytracing is out of the question, especially since the next gen will be gunning for 4K (most of the additional GPU power will end up here) . So we can expect sharper textures, better shadows and somewhat better GI on dynamic objects but I don’t think it’s going to be a major leap.
On the CPU side, however, it’s going to be a major step forward. Expect a Zen2 core on 7nm with 2 CCXs clocked at around 3.2-3.4 GHz, that is 8 cores running at twice the clock speed compared to PS4 with an IPC advantage of atleast 60%. That’s an increase of over 3 times compared to PS4 and that raw power is going to end up in much better physics and dynamics across the board. I’m especially excited about much better particle simulation in racing games and may be a physics based sound engine.
On the CPU side, however, it’s going to be a major step forward. Expect a Zen2 core on 7nm with 2 CCXs clocked at around 3.2-3.4 GHz, that is 8 cores running at twice the clock speed compared to PS4 with an IPC advantage of atleast 60%. That’s an increase of over 3 times compared to PS4 and that raw power is going to end up in much better physics and dynamics across the board. I’m especially excited about much better particle simulation in racing games and may be a physics based sound engine.
Some info from the slides. This should give a rough idea of what they're talking about. at certain points. I've tried to translate this into average-joe-level speak so don't hate if I've oversimplified anything.
P5 – Intro and presenter’s info
P10 – Lighting equation
P14 – p23 – Light accumulated from multiple points/directions. For the exterior HDRI a lot of light comes from all directions.
P24-28 – The Cornell Box is a common example. It has coloured walls and a single light at the top. In this case, also some faint ambient (flat) light. With one step, the light doesn’t carry colour around. The photon goes from the light, to the object, then rendered by the camera. With 3 steps, a photon goes from the light, picks up green from the right wall, bounced to the red and becomes redder, then onto the sphere. Note that it’s mostly rough surfaces (not glossy) so the light will bounce in many directions, not like a mirror or a laser, but you only have 3 bounces and limited number of rays which may produce odd results based on the RNG direction they get. This is one reason why you see a lot of grain in areas like under the sphere.
P47-50 – Existing useful guides and whitepapers.
P55 – Lightmaps! Basically, if you think of this as recording the lighting information to a texture. That way, when you apply the texture, it looks lit even without any lights being computed at the time.
P66-81 – Ambient Occlusion equations and methods. This is a sort of approximation that works backwards. Instead of lighting an object additionally using all the directions of a bright sky, you work out where the crevices of the object are and assume they can’t get lit from the entire sky, and are thus darker.
P82-83 – Spherical Harmonics. These work on octaves – in the example pic you might get 1 direction (blue all around), 2 directions (blue sky and yellow ground), 4 directions (nothing in the pic would look good, but you might use this as up/down/east/west to get the sky, ground, and both the lit and shadowed side of a sunset), etc etc etc. This is used as a cheap (fast to calculate, small storage size) source of ambient/environmental light.
P96-98 – GI. In the obviously-wrong picture, the light is calculated as a ray that goes straight to the outside environment map (hdri / panorama/ skybox). The surface under the car is bright road or lit grass, too bright so the underside looks too bright! With GI, the light has to come through the windows, and has to bounce (losing energy) multiple times before it can get into the footwells and crevices, etc.
P100-120 – some absurd baking numbers (Light calculation is called baking) even on an insane machine.
P128-132 – Look at the outcome with just per-vertex calculation (p131)! You cannot just use that, you have to use a corrected way with some type of per-pixel output (even if interpolated).
P139-143 – Trees are approximated as an ellipsoid when baking light (not planes/billboards/etc). Light can pass through the ‘leaf coverage’ which adds a green colour, and light that does bounce is scattered greatly because the leaves all point in different directions.
P148-152 – development WIP from early per-vert prototype to current PS4 state.
P168-170 – comparisons of the raytracing system on the PS4 VS the offline one (still MR in Maya?). Pretty good for baby steps, still the usual things like shadow resolution (due to different shadow calculations from realtime vs offline). Not sure why the Bugatti one looks so different, the tonemapper seems off with the raytracing example – see mirror and brighter surfaces?
P183-186 – how light bounces in many directions with rough surfaces. P183 is a really good example. A shiny surface on the left has most light reflected fairly evenly in mostly the same direction, the rough surface on the right scatters unevenly, and biases towards the surface normal (straight out/up).
P190-209 – Bounding Volume Heirachy. In simple terms, you divide up the (trees/lights/objects) into volumes based on visibility/importance. The goal is that you can work out what isn’t important and cull it so you’re not wasting processing time on unimportant things.
The rest isn’t particularly interesting or important. Note that they do have many quads on the Nissan GTR – good for tessellating while keeping things nice and smooth. The ground is tris and there are tris in the GTR too in some spots that are unavoidable (convergent point) or unimportant (window glass). Also, is that the paint chip car? I want that...
Some other questions – you could do dynamic day/night in GT Sport....If you had a system that could load both the 6pm and 8pm lighting data sets, and then average them for the output for 7pm and otherwise interpolate between them for 6:30pm, etc. If you want rain/clouds/etc, that’s another set of data you have to load and interpolate between. This is how GT6 did it, and why the 24hr versions didn’t look as good as the fixed-time versions of a track (fixed sized lighting data field divided by multiple sets for different times – allocated some time for calculations for blending between them = lower quality).
You CAN optimize the system – the ray doesn’t have to do all N bounces in the super-high-quality object! As each subsequent bounce adds less to the final result, you can use simpler models. This can hijack the LOD system much like the rear-view or reflection systems. You can also bias the ray towards the light source or the camera (viewer’s eye). This can improve speed for multiple bounces by trying to get an ‘early out’ case (either hitting a light or the eye). The sooner you can stop computing a ray, the faster the system will work. Does look bit different though.
Another thing you can do is not compute all rays each frame. Much like reflections only being solved at 30Hz or whatever, you can do bounces per frame – it will take (say) 3 frames to resolve, but as long as the lighting situation doesn’t change much in 3 frames you can get away with it. You can also do other blending with previous frames to reduce the noise/grain from using a limited number of rays.
Working out where to calculate lighting is a good way to reduce ray count. Will this patch of pixels have different lighting, so we have to do everything again, or will this be a smooth blend between these few points? This does go into it a bit in the slides; you can calculate a few points very accurately and the pixels in between are just a blend of several important points based on how close they are to them.
NOTE FOR PEOPLE LINKED HERE: Yes, this is me from the HD graphics page that linked you here.
P5 – Intro and presenter’s info
P10 – Lighting equation
P14 – p23 – Light accumulated from multiple points/directions. For the exterior HDRI a lot of light comes from all directions.
P24-28 – The Cornell Box is a common example. It has coloured walls and a single light at the top. In this case, also some faint ambient (flat) light. With one step, the light doesn’t carry colour around. The photon goes from the light, to the object, then rendered by the camera. With 3 steps, a photon goes from the light, picks up green from the right wall, bounced to the red and becomes redder, then onto the sphere. Note that it’s mostly rough surfaces (not glossy) so the light will bounce in many directions, not like a mirror or a laser, but you only have 3 bounces and limited number of rays which may produce odd results based on the RNG direction they get. This is one reason why you see a lot of grain in areas like under the sphere.
P47-50 – Existing useful guides and whitepapers.
P55 – Lightmaps! Basically, if you think of this as recording the lighting information to a texture. That way, when you apply the texture, it looks lit even without any lights being computed at the time.
P66-81 – Ambient Occlusion equations and methods. This is a sort of approximation that works backwards. Instead of lighting an object additionally using all the directions of a bright sky, you work out where the crevices of the object are and assume they can’t get lit from the entire sky, and are thus darker.
P82-83 – Spherical Harmonics. These work on octaves – in the example pic you might get 1 direction (blue all around), 2 directions (blue sky and yellow ground), 4 directions (nothing in the pic would look good, but you might use this as up/down/east/west to get the sky, ground, and both the lit and shadowed side of a sunset), etc etc etc. This is used as a cheap (fast to calculate, small storage size) source of ambient/environmental light.
P96-98 – GI. In the obviously-wrong picture, the light is calculated as a ray that goes straight to the outside environment map (hdri / panorama/ skybox). The surface under the car is bright road or lit grass, too bright so the underside looks too bright! With GI, the light has to come through the windows, and has to bounce (losing energy) multiple times before it can get into the footwells and crevices, etc.
P100-120 – some absurd baking numbers (Light calculation is called baking) even on an insane machine.
P128-132 – Look at the outcome with just per-vertex calculation (p131)! You cannot just use that, you have to use a corrected way with some type of per-pixel output (even if interpolated).
P139-143 – Trees are approximated as an ellipsoid when baking light (not planes/billboards/etc). Light can pass through the ‘leaf coverage’ which adds a green colour, and light that does bounce is scattered greatly because the leaves all point in different directions.
P148-152 – development WIP from early per-vert prototype to current PS4 state.
P168-170 – comparisons of the raytracing system on the PS4 VS the offline one (still MR in Maya?). Pretty good for baby steps, still the usual things like shadow resolution (due to different shadow calculations from realtime vs offline). Not sure why the Bugatti one looks so different, the tonemapper seems off with the raytracing example – see mirror and brighter surfaces?
P183-186 – how light bounces in many directions with rough surfaces. P183 is a really good example. A shiny surface on the left has most light reflected fairly evenly in mostly the same direction, the rough surface on the right scatters unevenly, and biases towards the surface normal (straight out/up).
P190-209 – Bounding Volume Heirachy. In simple terms, you divide up the (trees/lights/objects) into volumes based on visibility/importance. The goal is that you can work out what isn’t important and cull it so you’re not wasting processing time on unimportant things.
The rest isn’t particularly interesting or important. Note that they do have many quads on the Nissan GTR – good for tessellating while keeping things nice and smooth. The ground is tris and there are tris in the GTR too in some spots that are unavoidable (convergent point) or unimportant (window glass). Also, is that the paint chip car? I want that...
Some other questions – you could do dynamic day/night in GT Sport....If you had a system that could load both the 6pm and 8pm lighting data sets, and then average them for the output for 7pm and otherwise interpolate between them for 6:30pm, etc. If you want rain/clouds/etc, that’s another set of data you have to load and interpolate between. This is how GT6 did it, and why the 24hr versions didn’t look as good as the fixed-time versions of a track (fixed sized lighting data field divided by multiple sets for different times – allocated some time for calculations for blending between them = lower quality).
You CAN optimize the system – the ray doesn’t have to do all N bounces in the super-high-quality object! As each subsequent bounce adds less to the final result, you can use simpler models. This can hijack the LOD system much like the rear-view or reflection systems. You can also bias the ray towards the light source or the camera (viewer’s eye). This can improve speed for multiple bounces by trying to get an ‘early out’ case (either hitting a light or the eye). The sooner you can stop computing a ray, the faster the system will work. Does look bit different though.
Another thing you can do is not compute all rays each frame. Much like reflections only being solved at 30Hz or whatever, you can do bounces per frame – it will take (say) 3 frames to resolve, but as long as the lighting situation doesn’t change much in 3 frames you can get away with it. You can also do other blending with previous frames to reduce the noise/grain from using a limited number of rays.
Working out where to calculate lighting is a good way to reduce ray count. Will this patch of pixels have different lighting, so we have to do everything again, or will this be a smooth blend between these few points? This does go into it a bit in the slides; you can calculate a few points very accurately and the pixels in between are just a blend of several important points based on how close they are to them.
NOTE FOR PEOPLE LINKED HERE: Yes, this is me from the HD graphics page that linked you here.
Last edited:
- 3,279
- France
One of the best dev but Playground Games seems to be the new leader.Ray tracing on PS4 is insane. Need to see the details but that alone shows why PD are industry leading developers. The future is only going to better with the hardware bumps. PS5 should be 2020 so no doubt the next GT is already in the works
- 3,954
- Laguna
- Alpha_Cipher1
PD certainly has the potential to be an industry leading developer given just how much they can squeeze out of the hardware, but they can't be one if the hardware they're working with isn't industry leading anymore. It's not PD's fault, but industry leading developers and hardware goes both ways
- 14,431
- Los Angeles, CA
- Lma_robot
- lImaRobotl
Playstation needs it's own version of Horizon on it's side, it would be a hit. They can do it exactly the same way Horizon is handled. The cars are borrowed from the parent Forza Motorsport title, and Playground creates the gaming world they're used in.Imagine sandbox/open world racing game with great scenery like FH combined with detailed car model and epic lighting as GT. That one man can dream of...
- 982
Not really. This isn't a power race. Industry leading devs in tech are those who are able to come up with algorithms and systems within their box that produce outstanding results.
Several PS4 exclusives are considered industry leading: ND (modeling, animation, graphics), ICO (animation), PD (lighting, modeling)
- 3,279
- France
Imagine sandbox/open world racing game with great scenery like FH combined with detailed car model and epic lighting as GT. That one man can dream of...
Beautiful dream ^^. Now instead of focusing on tiny details or sim aspects, they should make the next GT about fun and enjoyment. Less FIA competition and more Mitsubishi FTO at Deep Forest listening to a rock song or DNB. Forza Horizon is the new system seller and I'm curious to know what Sony thinks about the racing genre on playstation.
- 2,596
- Brazil
Forza Horizon appeals to the same audience that fell in love with the Gran Turismo 1 back on the PS1 with a open world spin on it:
Realistic visuals
Handling model that is at the same time arcadey but good enough that it could pass as a entry level simracer for the casual eye
A great soundtrack
A big car list
- 4,538
- Australia
I think PD management places too much importance on the visual innovations than the game itself.
- 1,467
- Gent Ghent Gand
- svenvangent968
- 1,372
- Serbia
- Warlock__SRB
I think Kaz mentioned in one of his Q&A number of 250 or something... that is not small amount at all, although not all people are involved in game development... but still...
- 1,467
- Gent Ghent Gand
- svenvangent968
- 14,431
- Los Angeles, CA
- Lma_robot
- lImaRobotl
That estimate given by Kaz was also before they opened up a spot at Venice, California as well. So it's most likely somewhere between 200-300, since the article stated 200+ back then.Whaaaaaaaat 250 .........ok lets make it 200 mamamia .. to give you a idea Assetto Corsa worked at the beginning with 13 than the last time I heard something when they expanded was 30 persons .
- 919
- Philippines
200+ is actually quite small if you compare it to other AAA devs these days most of them have 400-1000 employees. Devs like Ubisoft, Rockstar, Bioware, Dice.Whaaaaaaaat 250 .....
Similar threads
