Game development presentations - a useful reference

[cheapchips]

@trinibwoy It will be interesting to see CDPR's use of UE5 for sure, but there's always the option that they basically don't use lumen and roll their own ray-tracing/GI. UE5 gives you a lot, but you have the option to fully customize any part of the engine that you want. They don't strike me as a company that will kind of roll with the out-of-box experience.

The announcement was declared as a partnership with Epic. While that could easily be interpreted as marketing guff, taken at face value it's more likely they'll contribute to Epic's solutions for existing features like Lumen, rather the branching off on their own.

 

[trinibwoy]

The entire selling point of ReSTIR/RTXDI is path traced direct lighting for primary views. If you can't use the algorithm in the presence of virtual geometry without introducing serious shadowing artifacts (poor acceleration structure) or high data structure cost then how much value do you think you're getting out of it ?

In the presentation they mention using Nanite for primary view rasterization. They also mention disabling screen space traces. So presumably ReSTIR DI is working well with Nanite in that particular demo. However they didn’t talk about the amount of geo in the BVH scene. It was a strange omission.

@trinibwoy It will be interesting to see CDPR's use of UE5 for sure, but there's always the option that they basically don't use lumen and roll their own ray-tracing/GI. UE5 gives you a lot, but you have the option to fully customize any part of the engine that you want. They don't strike me as a company that will kind of roll with the out-of-box experience.

Sure they can use all the UE5 animation, physics, particles, raster stuff and then roll their own lighting system. Based on the Nvidia talk hardware Lumen is doing a pretty good job though and can be further improved with a few tweaks - better light sampling algo and higher resolution probe grid.

 

[Lurkmass]

In the presentation they mention using Nanite for primary view rasterization. They also mention disabling screen space traces. So presumably ReSTIR DI is working well with Nanite in that particular demo. However they didn’t talk about the amount of geo in the BVH scene. It was a strange omission.

The only hard numbers that we got is that they were running 1080p/60hz w/ DLAA/RR on an RTX 4090 in an amusement park scene. We don't have much clear data about scene/acceleration structure complexity so it makes me think if they've introduced a ray traced shadow biasing mechanism or they're still using screen space shadows ?

 

[Frenetic Pony]

@trinibwoy It will be interesting to see CDPR's use of UE5 for sure, but there's always the option that they basically don't use lumen and roll their own ray-tracing/GI. UE5 gives you a lot, but you have the option to fully customize any part of the engine that you want. They don't strike me as a company that will kind of roll with the out-of-box experience.

They're using Lumen, they're using all the UE5 mainline features, which is why they switched in the first place. Nvidia's constant "RTX" version of UE5 is also pretty poor, so I don't know how this partnership will continue, especially since maxxing out UE5 already crushes GPU performance plenty on it's own.

As for restir, it's under R&D for UE5, it's a good technique, and the current concentration seems to be for direct raytraced shadows. The polycount for static objects is, potentially, solveable today. There's been a few good papers on how to manage streaming LOD that might work with some (Windows) API changes to allow for it on AMD, I don't know if Nvidia's hardware would support such though.

As for dynamic geometry, you can forget nanite skinned meshes/wpo being raytraced directly, it's the acceleration structure rebuild times that crush you. Because ray/box tests are relatively cheap, direct shadows are relatively coherent, and static meshes are static you can build a very tight acceleration structure around your triangles and you don't have a lot of divergence going through the BVH. Dynamic geometry throws much of that out.

 

[OlegSH]

There's been a few good papers on how to manage streaming LOD that might work with some (Windows) API changes to allow for it on AMD, I don't know if Nvidia's hardware would support such though

There's been papers on this from intel as well. Since BVH is fully built in sw, there is nothing stopping it from appearing on NVIDIA too, of course.

 

[trinibwoy]

As for dynamic geometry, you can forget nanite skinned meshes/wpo being raytraced directly, it's the acceleration structure rebuild times that crush you. Because ray/box tests are relatively cheap, direct shadows are relatively coherent, and static meshes are static you can build a very tight acceleration structure around your triangles and you don't have a lot of divergence going through the BVH. Dynamic geometry throws much of that out.

Nanite’s culling system will also need to account for wpo. And what happens to VSM caches? Dynamic geo is harder on everything not just RT.

It may come down to content. Open world UE5 games with lots of dynamic characters or animated foliage on screen may be a bridge too far on current hardware. Something like hellblade 2 may be doable though.

 

[Metricity]

Another one from GDC 2024:
Maximizing Graphics Performance with Flexible Virtualized Geometry
Supports the consoles and 3 different engines at the moment. Would like to see some fancier demos and comparisons with 1px triangles as they menioned that doesn't affect the quality. Also not sure about compression as the demo scene looks to be 50GB.

Virtualized geometry also looks to being worked on for the China version of Unity.