GPU Ray Tracing Performance Comparisons [2021-2022]

Maybe UE4 does not do SSR first and uses RT only for failure cases?
In that regard i wonder why this practice is a win at all. Maybe using RT always but only fetch the shaded material from screen where possible would be interesting to try.
UE4 4.27 and below have 3 RT reflection modes.
1. Default, used for Archvis and many games, high quality, very correct, just World Ray Traces
2. SSR + High Quality Archvis mode, same quality as (1) but blends SSR for increased performance
3. Experimental Deferred - lowest quality, most biased, has a built in scaler that makes lower res look better than mode 1 or 2.

Ascent is using 1. You can turn on 2 or 3 with Unreal Engine Unlocker and they genuinely increase performance by quite a bit... at the usual cost of correctness. The question is whether correctness really matters.
 
All you have to do is look at the Ascent's values for its RT in the dev console with Unreal Engine Unlocker to understand why its RT reflections are so expensive in comparison to other Unreal Engine 4 titles. It is using Full resolution reflections (most titles on the market are at best checkerboard usually (Fortnite), console games are quarter or 1/8th res. It has a very low roughness cut off (0.45 where other UE4 titles are often 0.1) and it includes transparency in reflections (particles).

Is it possible to change these parameters using the console, without breaking the game?
Some of those sound like a low-hanging fruit to get significantly higher performance with RT reflections with little fidelity degradation. Someone could create e.g. a 3rd party patch.

Or perhaps Neon Giant could enable those as options in the game settings with little developer intervention. That could even make it possible to enable RT on the Series X, with a resolution and framerate cost.
 
Is it possible to change these parameters using the console, without breaking the game?
Some of those sound like a low-hanging fruit to get significantly higher performance with RT reflections with little fidelity degradation. Someone could create e.g. a 3rd party patch.

Or perhaps Neon Giant could enable those as options in the game settings with little developer intervention. That could even make it possible to enable RT on the Series X, with a resolution and framerate cost.
I have not tried changing them for extended periods - but they are just visual quality options - I think they would not do anything to break it! One thing that could occur would be that loading new sections up might cause them to revert to the game designer default.
 
Are there not in game settings to reduce the resolution of RT reflections?
You can reduce them with Unreal Unlocker.

Did a few tests on 3090 (All max, Native res, DLSS Off).

4K RT Off for all effects - 94 FPS
4K RT Reflections On - 35 FPS, 10,41 ms for tracing rays
4K RT Reflections On Materials shading Off - 45 FPS, 6.35 ms for shading in hit points
4K RT AO On - 69 FPS - 2,67 ms for AO, BVH building - 1,175 ms
4K RT AO On, Reflections On - 32 FPS

The most obvious optimization would be reducing the number of rays by 4 times, this can be adjusted with the r.RayTracing.Reflections.ScreenPercentage in in-game console (need to unlock it first via the Unreal Unlocker).
With r.RayTracing.Reflections.ScreenPercentage set to 50, performance increases from 30 FPS to 48 FPS @ 4K, but reflections become too noisy on rough surfaces, so quality sucks, it's hard to blame devs for selecting full res.

Surprisingly, that's one of a few games that I tested, which benifits a lot from the hybrid RT/SSR reflections, so enbling them with the r.raytracing.reflections.hybrid 1 variable improves perf from 48 FPS to 52 FPS in the same scene.
These Hybrid reflections mess up quality in closed interior scenes, but sometimes they improve quality in open scenes, so I'd say they are neutral for image quality in general, but introduce some artefacting in the closed interior scenes.

The next optimization is the roughness threshold, as Alex said, this threshold is set on 0.45 by default in this game, so rays are cast for all surfaces that pass this threshold in a shaded pixel.
Surfaces with high roughness add the most of noise, so decreasing this threshold helps with noise.
Setting the roughness threshold to 0.1 with the r.RayTracing.Reflections.MaxRoughness command fixes the issues with lots of noise when r.RayTracing.Reflections.ScreenPercentage is set to 50, but graphics quality degrades due to decreased amount of reflections in scene (but still way better in comparison with SSR).
The roughness threshold opt futher increases perf from 52 to 56 FPS in 4K.

To improve image quality ever so slightly, one can use the r.RayTracing.Reflections.MaxRayDistance variable, when set to 1 000 000, it adds a little bit of additional shading from distant objects at mostly negligible cost of 0-1 FPS in scenes that I tested (this setting is likely highly scene dependent).

And most impactful optimization is the DLSS Quality mode, it doesn't degrade image quality in comparison with Native while frame rate grows from 56 to 82 FPS when other opts are enabled (you have to reapply r.RayTracing.Reflections.ScreenPercentage 50 after setting DLSS). At this point, perf becomes limited by my old 9900K CPU since GPU itilization drops from 99% to 91 - 93%. Frame rate is a bit lower in comparison with RT OFF's 86 FPS in the same scene, but image quality is still way better than RT Off and perf is mostly the same.

My settings of choice are default r.RayTracing.Reflections.ScreenPercentage with DLSS Quality, r.raytracing.reflections.hybrid 1, default 0.45 roughness threshold and r.RayTracing.Reflections.MaxRayDistance set to 1 000 000.
With such settings, perf is 60 FPS @ 4K (vs 50 FPS with default settings) and image quality is not affected if not better than default, so +10 FPS and better image quality in many cases in comparison with RT Reflections On and DLSS Quality at 4K, when compared to Native 4K, there is 2x difference without quality degradation.

RT variables can be seen here: https://developer.nvidia.com/blog/introduction-ray-tracing-unreal-engine-422/
Screens with different options are here: https://drive.google.com/drive/folders/11NAghOPtt6x206GezLiagorzRn49VSej?usp=sharing
 
Deathloop supports a very limited form of RT shadows, the 3090 is still 45% faster than 6900XT @1440p and 35% faster @2160p.
https://www.pcgameshardware.de/Deat...rk-Test-Review-Systemanforderungen-1379400/2/
Seems RT in the game can to some extent can be run on Pascal GPU's.
Older Radeon graphics cards lack a feature that Nvidia offers on the driver side - "software" ray tracing. As we were pleased to discover, ray tracing can be activated on a Pascal GPU (such as said Titan X, a GTX 1080 Ti or GTX 1060/6G) in Deathloop. The beam tracking is carried out purely by the shader ALUs. As expected, the performance costs are significantly higher than for the models with dedicated RT calculators (around 33 percent), but the radiation load is so low that even the GPU grandpa from 2016 still achieves well-playable frame rates - at least in Full HD.
 
Screenshot2021091520.png

https://www.computerbase.de/2021-09/deathloop-benchmark-review-test/
 
If it's not inevitable yet, it will become the norm soon. The real question that needs to be asked is whether we get to a point (within the next 3-4 years) where they stop benchmarking the non-RT versions of the games (because they can't). ie Metro Exodus.

I want us to get to that point. AMD cards may not be incredible at RT, but Metro Exodus shows that you can still have a sufficient amount of RT while cutting enough corners to make it still run sufficiently well.
 
With almost every important AA or AAA game now receiving RT implementations, I think it's ridiculous for any media outlet to do benchmarks without RT.

Which AAA/AA console games are launching without ray tracing?

Which mode? Does it count if it's only in Photo Mode?

If you mean by RT at all I could see your point (but still, there's room for all niches out there and non-RT games are far bigger than just niche), but if you mean they should only test RT stuff it's ridiculous.

Most people play games that are out now, not tomorrow. Steam current Top-9 most played games (it's top-10 but one isn't a game, not sure what Source SDK Base 2013 Multiplayer does on the list) have one game from 2021 and one from 2020, rest are older. And none of them has RT.
(and just so no straws should be pulled, I don't mean they should just test most played games either, it was just example of what people play)
 
If you mean by RT at all I could see your point (but still, there's room for all niches out there and non-RT games are far bigger than just niche), but if you mean they should only test RT stuff it's ridiculous.

Most people play games that are out now, not tomorrow. Steam current Top-9 most played games (it's top-10 but one isn't a game, not sure what Source SDK Base 2013 Multiplayer does on the list) have one game from 2021 and one from 2020, rest are older. And none of them has RT.
(and just so no straws should be pulled, I don't mean they should just test most played games either, it was just example of what people play)

I assume he’s referring to the common practice of disabling RT in reviews in games that support it. RT is still treated as an “extra” mode but we’re seeing more games coming out with RT enabled at normal/medium settings. So it would be silly for reviewers to continue with purist “RT disabled” defaults.
 
but if you mean they should only test RT stuff it's ridiculous.
Of course not, they should test a mixture of games like they do, but activate RT in those games that support it, as a part of running those games at Ultra/Max settings.

In any fair review right now, half of the tested games will include RT, games like: Metro Exodus, Deathloop, F1 2021, Dirt 5, Call Of Duty, Resident Evil, Cyberpunk, Fortnite, Crysis Remastered, Doom Eternal, Wolfenstein Young Blood, Battlefield V, Control, Watch Dogs Legion, Minecraft, Far Cry 6 .. etc, even Gears 5 should be tested with it's software ray traced Global Illumination, World of Tanks should also be tested with it's software ray traced shadows.

Reviewers should just test those games with RT enabled by default.
assume he’s referring to the common practice of disabling RT in reviews in games that support it
Yes exactly.
 
Last edited:
Back
Top