Ray/Path tracing and 6th Generation games. Possible on Current Gen Consoles?

[Dr. Nick]

After watching the Portal RTX DF video, I started thinking about something that I have not thought about much since the beginning of this generation: Are games from the PS2 generation (DC/PS2/GC/ OG Xbox) good candidates for the same treatment? Could they run at least 30 FPS on the PS5 and Series X? Is it possible with the series S?

Many games from the 6th generation still look good but have flat and drab lighting compared to even what is seen on Xbox 360 and PS3. Just like HD versions of games from the previous generation, I was looking forward to seeing RT versions of many of these games before the start of this generation. Then at the beginning of this generation, the consoles seemed like they just couldn't cut it at that kind of work.

Two years later, we have seen plenty of examples of these consoles doing ray tracing work that might not have seemed possible at the start of this gen. Not anywhere as good as what the RTX 2000/3000/4000 can do but still very impressive.
Until Portal (7th gen), most of the games that were getting the PT/RT Treatment were from what I believe would have been from somewhere between 5th gen/6th gen of gaming. Seeing how well it performs on the RTX 2060 using optimized settings seems encouraging.
So is there hope in seeing games of that generation getting similar path-traced treatment on consoles, even if it is lower quality than what you can get on PC?

 

[DavidGraham]

Minecraft RTX was shown to be running on Xbox Series X during the early days of promotions for the console, but it was just a marketing demonstration, and I think it ran 1080p30 with upscaling? Path Tracing qulaity was bared back too.

Anyways, I think it's possible with heavy upscaling.

 

[davis.anthony]

I think Quake 2 and Serious Sam have proven that older games absolutely benefit from path tracing.

Performance is the issue but older games that don't have that much high frequency detail typically upscale from lower resolutions better than newer games so it might be that much of a problem.

Silent Hill 2 path traced
Gran Turismo 4 path traced
MGS2 & 3 path traced

That would be amazing to see...

 

[techuse]

Fully path traced? I would say probably not at a resolution high enough to look very good.

 

[London Geezer]

Maybe I’m being silly but isn’t ray tracing better for reflections Compared to path?

 

[cwjs]

Maybe I’m being silly but isn’t ray tracing better for reflections Compared to path?

You may be mixing up ray marching (which due to the kind of data it runs on loses the precision necessary for accurate reflections) with path tracing -- path tracing is the same technique as ray tracing, just more of it: doing the entire render from ray casts and bounces, no traditional rendering.

 

[Inuhanyou]

I feel like it takes certain kind of game and geometry to make good RT. Slapping on a ps2 game without any other changes may look worse than the original art assets intended.

This is why RTX remix doesn't just slap RTGI on old games but attempts to actually clean up the assets too

 

[Cyan]

Minecraft RTX was shown to be running on Xbox Series X during the early days of promotions for the console, but it was just a marketing demonstration, and I think it ran 1080p30 with upscaling? Path Tracing qulaity was bared back too.

Anyways, I think it's possible with heavy upscaling.

maybe too heavy upscaling to be acceptable for console makers, at least if they go path traced. Quake II RTX is an excellent game to test the RT performance of a card, because it lets you choose a target framerate and the game decreases the resolution as much as needed to achieve that framerate.

 

[davis.anthony]

I feel like it takes certain kind of game and geometry to make good RT. Slapping on a ps2 game without any other changes may look worse than the original art assets intended.

Quake RT with the original textures and assets looks amazing though.

 

[Inuhanyou]

Quake RT with the original textures and assets looks amazing though.

It's pretty simple geometry game to be fair

 

[davis.anthony]

It's pretty simple geometry game to be fair

And more geometry will only make the results even better.

 

[Inuhanyou]

And more geometry will only make the results even better.

I mean ye. But...

 

[Shifty Geezer]

How does RT performance scale with geometry and shaders? Construction of the BVH will be faster with simpler geometry eliminating the overhead. I guess material shaders are pretty simple. So what factors slow down RT the most? If designing a game to be as RT efficient as possible, how would the assets be weighted?

 

[Dictator]

How does RT performance scale with geometry and shaders? Construction of the BVH will be faster with simpler geometry eliminating the overhead. I guess material shaders are pretty simple. So what factors slow down RT the most? If designing a game to be as RT efficient as possible, how would the assets be weighted?

avoid long and thin geometry, avoid using alpha tested geometry where possible, merge geometries and instances to perhaps increase GPU utilisation...

There are probably a lot of them!

 

[davis.anthony]

@OlegSH can we have your input on Shiftys question?

 

[Shifty Geezer]

avoid long and thin geometry, avoid using alpha tested geometry where possible, merge geometries and instances to perhaps increase GPU utilisation...

There are probably a lot of them!

What's the relationship between geom complexity and RT demands? Is BVH construction O, O(logn), O(n^2)? Likewise BVH traversal? And is there a sweet-spot where the two are nicely balanced?

 

[cwjs]

What's the relationship between geom complexity and RT demands? Is BVH construction O, O(logn), O(n^2)? Likewise BVH traversal? And is there a sweet-spot where the two are nicely balanced?

BVH construction is n squared, traversal is log n, but the number of traversals needed varies a lot by surface complexity and bounding volume quality. For example, long triangles are expensive because it's highly likely that given a ray will pass close enough to one to require traversing to the leaf it's in to gaurantee it's a miss. Here's a good resource: https://www.pbr-book.org/3ed-2018/P...tion_Acceleration/Bounding_Volume_Hierarchies.

Seeing as these are all gpu operations that have to run approximately a gazillion times per frame, the real performance concerns probably have more to do with cache coherency and synchronizing all of the parellel work than they do algorithmic complexity.

 

[trinibwoy]

What's the relationship between geom complexity and RT demands? Is BVH construction O, O(logn), O(n^2)? Likewise BVH traversal? And is there a sweet-spot where the two are nicely balanced?

BVH construction should on average be relatively cheap even with increased geometric detail assuming you’re not rebuilding every BLAS every frame. I don’t think TLAS builds care about individual triangles.

Traversal slows down because they are lots more boxes to intersect and also more cache thrashing as a result. Even with more fine grained boxes leaf nodes probably end up with more geometry anyway so triangle intersection slows down too.

It’s unfortunate we still don’t have synthetic benches for this stuff 4 years after DXR launch. It’s like not having fillrate benches back in the day.

 

[OlegSH]

can we have your input on Shiftys question?

In addition to what has been said in other posts, while geometry topology should be considered to improve performance with RT, I don't think that's a problem in current titles and it will be even less of a problem in future ones.
Photogrammetry, UE5 and automated content generation drive industry towards more polygon soup like content, so it becomes more problematic and expensive to moderate it manually.
With microtriangles, it's assumed that LOD systems, authoring tools, editors and better geometry compression schemes should take care of the content complexity so that no manual optimizations are required. This means better geometry management tools and better geometry compression will be in scope in the upcoming years. Without better compression schemes, RT content will hit storage wall long before it will hit traversal limitations with Nanite levels of geometry. Also, traversal scales logarithmically vs linear scalign of BVH building, so there are already schemes that trade traversal time for faster BVH construction and lower memory footprints, such as Ada's DMMs that allow to compress geometry in a watertight manner without inquiring additional BVH construction or vertex shading cost, + they also speed up any-hit shaders by encoding state in micromeshes.

 

[OlegSH]

Practical Tips for Optimizing Ray Tracing | NVIDIA Technical Blog

To achieve high efficiency with ray tracing, you must build a pipeline that scales well at every stage. This starts from mesh instance selection and their data processing towards optimized tracing and…

developer.nvidia.com