The "optimization," is just reducing the number of light bounces, which degrades quality.
People always talk about games being "poorly optimized," but don't actually know what that means. Sometimes a game is just too demanding for current hardware. That used to be a thing before we demanded 4K 120 Ultra on day 1.
This mod that takes down the bounce count still gives a better experience than standard RT if you're a fidelity snob and not a frame rate snob. It's pretty neat to see it running playably at 1080p on a 3060, which is now the most installed GPU according to steam.
Default is 2 bounces / 2 rays per pixel.
I believe the "mod" setting they're talking about in the video is 1 bounce / 1 ray per pixel (with the 1 bounce being why reflections don't look great). Still, definitely makes it more generally usable for a lot of people.
After playing with it today, I have found that the 4090 makes 2 bounces / 4 rays per pixel pretty dang usable and **very** good looking. Native res of 3440x1440, DLSS Balanced and Frame Gen, ~90ish fps.
Some people may scoff at that internal res being run on a 4090, but the combination of how many rays are being shot out per pixel and the enhancements to DLSS make it a much more stable image than I would have expected. And this use case has also personally sold me on frame gen.
The mod is an edited:
>base\weather\24h_basic\cp2077_master_env_nge_v002.env
where:
bounceNumber = 1
rayNumber = 2
it also originally changed rangeFar = 2000 to 1600 but this doesn't seem to change anything? It was reverted back to 2000 anyway. More research needed to figure out what some of these parameters do.
The optional part of the mod is an .ini file that enables RIS (Resampled Importance Sampling), which is a developer toggle.
You can set dev toggles in real-time via Cyber Engine Tweaks console:
>GameOptions.SetBool("RayTracing/Reference", "EnableRIS", true)
It is false by default. exe strings were dumped on day one and there are *a lot* of new hidden options for the path tracer and denoiser. You can grab the list from the modding discord and play around with them in game using:
>GameOptions.SetBool("", "", value)
Depending on the value type, you will use SetFloat (for decimal numbers), SentInt (for integer numbers) or SetBool (for true/false statements)
RIS is a general technique and I'm almost certain Cyberpunk uses it by default, so it's a bit hard to tell what EnableRIS means, but I'd stare at https://github.com/NVIDIAGameWorks/RTXDI/blob/main/doc/Integration.md to find out.
Thanks for this link. The nVidia whitepaper on RIS is extremely math heavy and mostly incomprehensible to me so I really have no idea how it works.
And Cyberpunk hidden GameOptions don't always have names that make sense, like the aforementioned alpha shifts. Complete moonshot guess: whatever light scattering model they used had some constant called (greek letter) alpha or something.
Its pretty easy. Go get Wolvenkit Nightly: https://github.com/WolvenKit/WolvenKit-nightly-releases
Add base\weather\24h_basic\cp2077_master_env_nge_v002.env to your project and convert it to .json. Now you can edit/find/replace in notepad++ or vscode.
When you are done, convert json to env and click the install button and you are good to go.
From what I have read photo mode is at 6 bounces 2 Rays by defualt.
that could explain why it gets higher quality like this
[https://twitter.com/Berduu/status/1645900112344891394](https://twitter.com/Berduu/status/1645900112344891394)
yeah even with the lower bounce count, its still path tracing on friggin cyberpunk! the fact that even low end gpu can actually play cyberpunk path tracing at all is amazing and for them well worth the slight visual difference
I might try that too, just started with 2B4R.
And in loading it up again, I was on the low side in terms of frame rate. A less demanding area is around 120.
> That used to be a thing before we demanded 4K 120 Ultra on day 1.
Kind of nonsense when the complaining is on lower end hardware not being able to run said games.
I mean, it's just a setting that should be in the in game menu and I'm sure CDPR will implement it at some point. I presume it wasn't there pun launch because of the Nvidia partnership, as marketing RT Overdrive as something that can only be ran by the new RTX 4000 series would surely make some people get a new card just for this.
As the initial hype dies down and people realize you can tweak it to work with older cards there's no reason to not let people do this from the game itself.
They may do in future although it would have to be in a more user friendly way. Hidden GameOptions are hidden for a reason: https://pastebin.com/TjAnnjg0
Thats nearly 1,500 parameters. Many of them are related and affect each other so setting the value of one parameter may have no effect if the value of another parameter doesn't make any sense.
I had some fun with the mod.
I can get it to around locked 30fps with XeSS Performance mode on my 6700XT. It obv doesn't look that great. Better than whatever FSR is trying to do though.
> The CyberFSR mod was so superior that its not even funny. And thats made by 1 person tinkering with stuff.
If I had a nickel for every time a single modder fixed a major bug, I might be able to buy "midrange" GPU.
https://www.rockpapershotgun.com/a-smoother-skyrim-for-christmas
> Dec. 23, 2011
> The creator reckons it can add a performance boost of around 40% in CPU-dependent scenarios; for instance, around 10 frames per second to chug hotspots such as the lovely waterfall-based city of Markarth. It's been achieved, apparently, by fixing an alleged tiny oversight on the part of Bethesda.
> Claims unofficial path-maker Arisu, "It works mostly by rewriting some x87 FPU code and inlining a whole ton of useless getter functions along the critical paths because the developers at Bethesda, for some reason, compiled the game without using any of the optimization flags for release builds." Also, "They'd just have to add the string "/O2 /GL" to their Makefile". That easy, huh?
I remember Skyrim at launch would drop below 30 FPS even on a ~5 GHz i5-2500K from running x87 instructions for floating point calculations.
I suspect they had some bugs and decided the easiest way to deal with them was to disable all compiler optimizations instead of actually tracking down the root causes, in order to launch the game “on schedule”.
There was a lot of stuff stripped from the game in order to make the release date. For example, there was going to be a huge living civil war throughout Skyrim but it was axed. A modder found the unimplemented files in the code and he made a mod that more or less enabled it
>Whoever at CDPR implemented FSR 2.1 into Cyberpunk needs to be ashamed of themselves.
Compared to CyberFSR2 mod it's *almost* like NVIDIA paid them to make the official FSR2 implementation look worse.
Its honestly beyond me how bad the image stability and ghosting is in the official implementation. I can't believe someone took a look at that and said. Yep this is the one.
And its not a isolated issue for CDPR. Their Witcher 3 implementation is also terrible and also gets beaten by the mod.
right? the fsr in this game is so very bad. and we know it can be better. just look at well most the games hardware unboxed featured recently in their video.
man i can believe how much more expensive that extra bounce is! id have thought it woudlnt be near that much more over just one bounce. i am loving the mod thou, yeah some specular surfaces suffer a bit but i can play 1440p dlls quality path tracing cyberpunk at >40fps on my 3090. and with my freesync screen its really nice.
The more I look at results from the RT Overdrive mode the more confused I get.
The 3070 and 2080Ti have always been pretty much a match in both raster and RT performance. But here the 2080Ti falls below the 3060.
Then there's another big outlier, the A770. DF didn't test it, but the RT hardware on the A770 is supposed to outclass that of competing RDNA2 and Ampere GPUs and is more in line with Lovelace. And while the A770 leads over RDNA2 here, [it still falls way short of the 3050.](https://cdn.discordapp.com/attachments/1017221825996005478/1097230744599920730/image.png)
While it's certainly impressive to see path tracing in a modern game, this just feels lile a tech demo specifically optimised to run solely on 30 and 40 series GPUs, and that's it.
What do you mean? In his chart it shows the 2080ti being 20% faster than the 3060.
I think there are optimizations in the 3000 series that probably just lend themselves better to path tracing, and heavy workloads. things that lend itself better to path tracing. Optimization's that just don't show when it's a very light workload.
Those 'Ampere optimizations' are raw compute and accelerated RT motion blur.
The reason why Ampere had 'fake TFLOPs' is that according to Nvidia, RT frames are 50% composed of compute ops. Also, tensor ops are part of it. So higher TFLOPs and faster tensor cores means better path tracing.
The same whitepaper says that 36% of ops in games are integer, which explains the poor TFLOP/FPS scaling of Ampere in raster compared to pathtracing in cyberpunk or blender
> Then there's another big outlier, the A770. DF didn't test it, but the RT hardware on the A770 is supposed to outclass that of competing RDNA2 and Ampere GPUs
He said that its a slideshow and needs a driver update from Intel.
It's almost as bad as RDNA2 is faring, surely you'd assume the same of RDNA2? Especially considering that AMD haven't released an optimised driver for this release yet.
6900XT was usually close to 3070's level in the newer PT games, instead of being half of its performance. And you can see 2080Ti trailing 3070 significantly just like here.
https://www.pcgameshardware.de/Doom-Classic-Spiel-55785/Specials/Raytracing-Mod-PrBoom-Benchmarks-1393797/2/#a2
https://www.pcgameshardware.de/Serious-Sam-The-First-Encounter-Spiel-32399/Specials/SeSam-Ray-Traced-Benchmark-Test-1396778/2/
Even in quake II RTX, 6800XT was half of 3080's performance, before driver optimizations and not just 40% of 3070 like here.
https://www.guru3d.com/news-story/radeon-rx-6800-xt-performs-half-as-good-compared-to-rtx-3080-in-vulkan-ray-tracing-tests,3.html
> It's almost as bad as RDNA2 is faring, surely you'd assume the same of RDNA2?
I don't to be honest , ARC has some good dedicated RT accelerators , while RDNA2 doesn't.
Both have dedicated accelerators, ARC just also has BVH Traversal in HW. There's been no info how many intersection tests ARC can do though. Considering the shader executors and the raytracing units communicate via L1 on ARC though, the communication is probably even faster on RDNA2.
(That's not to say that AMDs RT performance has been great, it hasn't).
48 Comments
Substance___P@reddit
From-UoM@reddit
mac404@reddit
Pokiehat@reddit
Veedrac@reddit
Pokiehat@reddit
pieking8001@reddit
Pokiehat@reddit
mac404@reddit
From-UoM@reddit
edernlg@reddit
mac404@reddit
pieking8001@reddit
Qesa@reddit
mac404@reddit
bazooka_penguin@reddit
kingwhocares@reddit
cremvursti@reddit
Pokiehat@reddit
conquer69@reddit
TheSilentSeeker@reddit
Firefox72@reddit
COMPUTER1313@reddit
Atemu12@reddit
COMPUTER1313@reddit
specter491@reddit
rubenssm@reddit
HandofWinter@reddit
b3081a@reddit
pieking8001@reddit
Firefox72@reddit
pieking8001@reddit
Firefox72@reddit
pieking8001@reddit
pieking8001@reddit
uzzi38@reddit
bubblesort33@reddit
ResponsibleJudge3172@reddit
theoutsider95@reddit
uzzi38@reddit
dparks1234@reddit
bctoy@reddit
theoutsider95@reddit
L3tum@reddit
OwlProper1145@reddit
OwlProper1145@reddit
Firefox72@reddit
panix199@reddit