What do gamers actually think about Ray-Tracing?

[Rroff]

Optimizing Mega Lights is one thing, using a texture so you avoid using actual geometry for a relative "flat" floor only "optimizes" for that scene (assuming you don't have other complex assets in the scene). Once you go out, you encounter vegetation (also Nanite), possible plenty of characters (also nanite), perhaps snow/mud (also dynamic nanite tesselattion) to which, if you apply the same "optimization" pass, you'd need to make them "old gen".

As for PT, as I've said, it runs fine on my 4080... Even a 4070 does decently.

Properly done path tracing shouldn't be affected too much by scene complexity, until you add in more advanced stuff like caustics, Unreal Engine seems to have made a bit of a meal of it with one foot stuck in software Lumen, etc.

 

[Tinek]

Optimizing Mega Lights is one thing, using a texture so you avoid using actual geometry for a relative "flat" floor only "optimizes" for that scene (assuming you don't have other complex assets in the scene). Once you go out, you encounter vegetation (also Nanite), possible plenty of characters (also nanite), perhaps snow/mud (also dynamic nanite tesselattion) to which, if you apply the same "optimization" pass, you'd need to make them "old gen".

You don't need Nanite for every single asset, unless you want FPS to suffer. Nanite is great for cases where you want to add to the game super complex, high-poly assets - it speed things up then. But if you have a flat surface and slap nanite on it, you kill FPS for nothing. If you have low details on the object visible to player, but underneath super high poly count - again, you kill FPS for nothing. Devs need to run mesh optimisation path beforehand and not just take every single item straight from for example Blender and put it inside game without any consideration to how it influences FPS. Yes, it will be faster with Nanite than without it, but it will be much slower with Nanite than with just optimised mesh. Nanite has its use, but just turning it on and forgetting about any mesh optimisation will result in much lower FPS without any visual advantage to the gamer. And surprise - that's what happens in modern UE games a lot, which I call lazy development.

As for PT, as I've said, it runs fine on my 4080... Even a 4070 does decently.

Define fine. On 4090 (full power limit, to that 7950x3D), in I.J. with latest patches (they did improve performance a bit for me), I get in the jungle around 40FPS with full PT, 1440p UWQHD resolution, DLSS Quality (that's only 960p!). That's without FG. May I remind you that even NVIDIA advises strongly to achieve at least 55FPS before enabling FG, to not feel excessive visual lag? Personally I start to feel sick quite quickly with it being below 60 of input FPS, as latency between my input and image reaction is just too easy to feel. This is also consistent with what DF said in their videos (around 60FPS of input to feel good). That said, in I.J. FG for me is still broken, causes horrible stutter which makes game absolutely unplayable (not just by feeling, it really is just unplayable, as even aiming camera at anything is near impossible).

That said, for PT to have good quality and as little noise as possible, good reflections quality (resolution and by that details), etc. you need to have enough pixels for the denoiser to work with without removing too many details. 960p is just not enough with so small number of rays. DLSS sharpens image and does nice AA (especially on thin lines) but there's nothing in it that could combat noise or other issues causes by low resolution of PT rendering, by the very physical nature of how PT work. You need different kind of AI for that - the ones that NVIDIA will likely introduce in Jan, but these add latency and lower FPS too. All these compromises and it's still just 40FPS in my case. I can see with FG that would be around 80-100FPS (if it worked properly in this game for me) but that has its own issues with added latency, as mentioned earlier. That, in my definition, is not fine running. I can't imagine how 4080 could do it better than much faster 4090.

 

[Tinek]

Properly done path tracing shouldn't be affected too much by scene complexity, until you add in more advanced stuff like caustics, Unreal Engine seems to have made a bit of a meal of it with one foot stuck in software Lumen, etc.

UE doesn't use PT in hardware Lumen path, does it? AFAIK, they have their own approach to RT that isn't just pure PT but in effect it has the chance to run faster and with less issues, depending on the skill of devs implementing it. They also pretty much ditched support for software Lumen with 5.5 version, as per Mega Lights dev (it's still there, but not being developed anymore, apparently).
That aside, real PT caustic is too slow even for render farms in Hollywood, hence they almost always fake it with animated textures etc. and you mention it in live rendering as if that could happen...

 

[Rroff]

That aside, real PT caustic is too slow even for render farms in Hollywood, hence they almost always fake it with animated textures etc. and you mention it in live rendering as if that could happen...

Fast approximation is possible (with reasonable results) - Quake 2 RTX does a very basic implementation of a fast approximation version of caustics but that is mostly because it isn't used much/at all by the original game levels, rather than performance reasons, so they didn't spend time on a feature which wasn't really required unfortunately.

Unfortunately most implementations of ray tracing/path tracing outside of Quake 2 RTX are still one foot in hybrid techniques, which sap performance over a pure implementation, with many still using a low quality version of the world to process all or some features and then paint into the higher detail world with less than ideal results.

 

[Tinek]

Unfortunately most implementations of ray tracing/path tracing outside of Quake 2 RTX are still one foot in hybrid techniques, which sap performance over a pure implementation, with many still using a low quality version of the world to process all or some features and then paint into the higher detail world with less than ideal results.

And the reason for that is what I mention as the main problem - our current best GPUs are orders of magnitude too slow to do actual proper real PT fast enough to have it implemented in games like that. As AMD said in their material about speeding up RT processing, hundreds of TB/s memory is needed for RT cores to get proper speed, which currently is impossible, as even GDDR 7 is many orders of magnitude too slow. Their idea is to add pretty much super fast L1 cache to every small RT cores cluster but even that is far too slow, unless one can connect them into big clusters and use math trickery to "simulate" fast enough memory for the whole frame. That will happen, apparently, in future GPUs of their design, not the upcoming 9000 series. I reckon NVIDIA has similar ideas for the future. We're just not there yet, for years to come.

That said, Q2 is full of noise, low resolution reflections etc. too - they use heavy denoising on the image, as even such simple/old game is way too much to handle with proper quality by modern fastest GPUs.

 

[Rroff]

That said, Q2 is full of noise, low resolution reflections etc. too - they use heavy denoising on the image, as even such simple/old game is way too much to handle with proper quality by modern fastest GPUs.

Full resolution reflections in Quake 2 RTX (mostly) the path tracer in the game isn't unduly troubled by increased detail or benefiting that much from being an older game as it is still running the full feature set (as far as it goes) regardless of what it is rendering - the only thing which slows it down is rendering complex vegetation, etc. and that is mostly due to the original game engine having limited handling of alpha blending/transparency rather than a path tracer problem as such. Noise is the bigger issue but not unsurmountable - people have been playing about with stuff like Area-ReSTIR and there are reconstruction techniques which the game doesn't use which can vastly improve it. Anything from the 4080 upwards could also handle a lot higher ray count than the renderer uses by default while still giving playable performance so as to reduce it as well - though we need something more like twice the 4090 performance to really put it in a place where noise would be barely an issue.

 

[Calin Banc]

You don't need Nanite for every single asset, unless you want FPS to suffer. Nanite is great for cases where you want to add to the game super complex, high-poly assets - it speed things up then. But if you have a flat surface and slap nanite on it, you kill FPS for nothing. If you have low details on the object visible to player, but underneath super high poly count - again, you kill FPS for nothing. Devs need to run mesh optimisation path beforehand and not just take every single item straight from for example Blender and put it inside game without any consideration to how it influences FPS. Yes, it will be faster with Nanite than without it, but it will be much slower with Nanite than with just optimised mesh. Nanite has its use, but just turning it on and forgetting about any mesh optimisation will result in much lower FPS without any visual advantage to the gamer. And surprise - that's what happens in modern UE games a lot, which I call lazy development.

Define fine. On 4090 (full power limit, to that 7950x3D), in I.J. with latest patches (they did improve performance a bit for me), I get in the jungle around 40FPS with full PT, 1440p UWQHD resolution, DLSS Quality (that's only 960p!). That's without FG. May I remind you that even NVIDIA advises strongly to achieve at least 55FPS before enabling FG, to not feel excessive visual lag? Personally I start to feel sick quite quickly with it being below 60 of input FPS, as latency between my input and image reaction is just too easy to feel. This is also consistent with what DF said in their videos (around 60FPS of input to feel good). That said, in I.J. FG for me is still broken, causes horrible stutter which makes game absolutely unplayable (not just by feeling, it really is just unplayable, as even aiming camera at anything is near impossible).

That said, for PT to have good quality and as little noise as possible, good reflections quality (resolution and by that details), etc. you need to have enough pixels for the denoiser to work with without removing too many details. 960p is just not enough with so small number of rays. DLSS sharpens image and does nice AA (especially on thin lines) but there's nothing in it that could combat noise or other issues causes by low resolution of PT rendering, by the very physical nature of how PT work. You need different kind of AI for that - the ones that NVIDIA will likely introduce in Jan, but these add latency and lower FPS too. All these compromises and it's still just 40FPS in my case. I can see with FG that would be around 80-100FPS (if it worked properly in this game for me) but that has its own issues with added latency, as mentioned earlier. That, in my definition, is not fine running. I can't imagine how 4080 could do it better than much faster 4090.

Let's put it this way... if a flat floor poses a problem, again, what you do with nanite on vegetation, landscape, perhaps not-so-flat-floors (dynamic nanite tesselattion) or characters? You turn everything off?

In the opening section, with jungle, 4k DLSS Performance (1080p), with usual stuff off (motion blur, dof, etc), FoV around 105 I think, I get about high 40s to 50s and 50s towards 60s with Vegetation Dynamic Animation off (or whatever is that called). Either way, is a non-issue since is mostly a walking simulator at that point, stutter free and pretty smooth with or without FG. A bit laggy, but still... irrelevant for most people. If I go Ultra Performance jumps to 60fps I think. Not too bad, since should be a lot better on a 4090 which is more of a 4k card than a 4080. I doubt the top dog in Crysis days could max it out at that time highest resoltuion and pull 60fps non stop.

That's with latest DLSS 3.8.10.0, updated via DLSS Swapper.

The question would be: how much more performance you can sqeeze out if you ditch completely the old lagacy "classic raster" and optimize heavily towards the new? Probably an issue would be with AMD since it tends to trip on heavy RT/PT.

 

[Tinek]

Full resolution reflections in Quake 2 RTX (mostly) the path tracer in the game isn't unduly troubled by increased detail or benefiting that much from being an older game as it is still running the full feature set (as far as it goes) regardless of what it is rendering - the only thing which slows it down is rendering complex vegetation, etc.

There's no complex vegetation in Quake 2, though? As in, there's no vegetation at all, it's 99% indoor game.

Noise is the bigger issue but not unsurmountable - people have been playing about with stuff like Area-ReSTIR and there are reconstruction techniques which the game doesn't use which can vastly improve it.

And yet we've seen no sight of any of that in other games with PT either? Unless you count in RR, which has issues too and doesn't remove the problem fully either (even best AI won't help with just too low resolution and too much noise). MegaLights dev mentioned one has to be very careful designing games with RT to not overwhelm denoiser as is and that's not even full PT either. One would hope new AI based stuff will help but it cost performance too, so - as usual - it will have to be a balance between performance and quality.

Anything from the 4080 upwards could also handle a lot higher ray count than the renderer uses by default while still giving playable performance so as to reduce it as well (...)

A lot higher ray count means what? 4 instead of 2? We currently see about 4-5 max ray counts per pixel in best current RT and PT implementations. We need orders of magnitude more to have proper quality without AI dreaming up missing things. Sure, devs could actually do their job and optimise stuff much better than they currently do (on average) and maybe then we could get a few more rays or xx60 class GPU could handle it ok with current quality. But no, GPUs can't handle higher counts at all (not even 4090) - I have described earlier what AMD said is needed to combat that and it's not about transistors in the GPU anymore (or GPU performance in general). It's about memory speeds apparently. Especially if you want to connect it with AI processing - GDDR 7 is orders of magnitude too slow to handle that, so is L1 cache in itself apparently. New clever design of GPUs is needed to improve things.

 

[Tinek]

Let's put it this way... if a flat floor poses a problem, again, what you do with nanite on vegetation, landscape, perhaps not-so-flat-floors (dynamic nanite tesselattion) or characters? You turn everything off?

You're moving in circles, I already responded to this.

In the opening section, with jungle, 4k DLSS Performance (1080p), with usual stuff off (motion blur, dof, etc), FoV around 105 I think, I get about high 40s to 50s and 50s towards 60s with Vegetation Dynamic Animation off (or whatever is that called). Either way, is a non-issue since is mostly a walking simulator at that point, stutter free and pretty smooth with or without FG. A bit laggy, but still... irrelevant for most people. If I go Ultra Performance jumps to 60fps I think. Not too bad, since should be a lot better on a 4090 which is more of a 4k card than a 4080. I doubt the top dog in Crysis days could max it out at that time highest resoltuion and pull 60fps non stop.

That's with latest DLSS 3.8.10.0, updated via DLSS Swapper.

I have finally figured out what was killing my FG and the FPS in game (as mentioned earlier, my CPU usage was higher than GPU one) - NVIDIA drivers settings (Ultra low latency turned on). Factory reset to all drivers settings and it's fine again. Now I get low CPU usage, 100% GPU usage, FG working (though it's unusably laggy anyway in this game, so I can't use it - it's way too floaty for some reason, with over a 100ms latency it feels). PT, DLSS Quality (960p) and I get about 70-80FPS, no FG. That's usable. Problem with DLSS Quality is that vegetation (leaves) have huge amount of white noise in animation (likely too low input resolution overwhelming denoiser), which isn't the case with just TAA turned on and no DLSS, but then it's below 50FPS (80 with FG but again, very floaty/laggy with what feels to be over 100ms latency in camera movement). Still, ignoring vegetation, all else looks fine so far. Oddly, in CP2077 and other games FG feels much better than in this one, not sure what they did wrong with implementation, it's worse than triple buffered vsync... :/ And yes, I have newest DLSS updated with DLSS Swapper always, in all games.

The question would be: how much more performance you can sqeeze out if you ditch completely the old lagacy "classic raster" and optimize heavily towards the new?

With current hardware and types of games - if one follows actual ML advise, like TI guy shown, in specific scenes one can run such things on 3060 with sensible FPS. I do recall playing CP2077 RT on 3060Ti and having over 90FPS with FG, was very playable in 1440p. That's hybrid approach though. If you ditch ALL raster and switch to 100% PT, you won't gain any performance, you will suddenly have a lot of CUDA cores doing nothing, RT cores fully saturated, GPU semi-bored yet FPS dropping. Current GPUs are NOT designed for full PT, it's simple as that. And (as mentioned earlier) you can't make them much faster without full redesign, clever cache memory assigned to RT cores, etc. apparently. UE5.5 doesn't seem to be full PT, though, so should be more suitable for current design of GPUs - but again, devs have to optimise things more than they currently are doing.

Probably an issue would be with AMD since it tends to trip on heavy RT/PT.

AMD nearly doesn't even exist on markets' graphs, they have 0 influence on PC gaming market. And on consoles they partnered with Sony to do that exactly - push RT much further (as mentioned earlier) but that takes time and won't come soon.

 

[tommybhoy]

As for PT, as I've said, it runs fine on my 4080... Even a 4070 does decently.

Even a 4070 does decently.

Watching a capped 60fps YT vid doesn't=decent gameplay.

66-78 fps DLSS3, I'd NEVER run that fps on FG, using your example, my 4070 can technically do PT.

 

[Calin Banc]

You're moving in circles, I already responded to this.


I have finally figured out what was killing my FG and the FPS in game (as mentioned earlier, my CPU usage was higher than GPU one) - NVIDIA drivers settings (Ultra low latency turned on). Factory reset to all drivers settings and it's fine again. Now I get low CPU usage, 100% GPU usage, FG working (though it's unusably laggy anyway in this game, so I can't use it - it's way too floaty for some reason, with over a 100ms latency it feels). PT, DLSS Quality (960p) and I get about 70-80FPS, no FG. That's usable. Problem with DLSS Quality is that vegetation (leaves) have huge amount of white noise in animation (likely too low input resolution overwhelming denoiser), which isn't the case with just TAA turned on and no DLSS, but then it's below 50FPS (80 with FG but again, very floaty/laggy with what feels to be over 100ms latency in camera movement). Still, ignoring vegetation, all else looks fine so far. Oddly, in CP2077 and other games FG feels much better than in this one, not sure what they did wrong with implementation, it's worse than triple buffered vsync... :/ And yes, I have newest DLSS updated with DLSS Swapper always, in all games.


With current hardware and types of games - if one follows actual ML advise, like TI guy shown, in specific scenes one can run such things on 3060 with sensible FPS. I do recall playing CP2077 RT on 3060Ti and having over 90FPS with FG, was very playable in 1440p. That's hybrid approach though. If you ditch ALL raster and switch to 100% PT, you won't gain any performance, you will suddenly have a lot of CUDA cores doing nothing, RT cores fully saturated, GPU semi-bored yet FPS dropping. Current GPUs are NOT designed for full PT, it's simple as that. And (as mentioned earlier) you can't make them much faster without full redesign, clever cache memory assigned to RT cores, etc. apparently. UE5.5 doesn't seem to be full PT, though, so should be more suitable for current design of GPUs - but again, devs have to optimise things more than they currently are doing.


AMD nearly doesn't even exist on markets' graphs, they have 0 influence on PC gaming market. And on consoles they partnered with Sony to do that exactly - push RT much further (as mentioned earlier) but that takes time and won't come soon.

Arguing for "selective use" of nanite is pointless, you either go all in or not all - perhaps better to just stick cu regular tessellation then. I really don't care for such optimization which is basically staying with the old, withhold progress. Like I've said, to me it's either full or nothing. TI isn't optimizing like that, he's getting firmly stuck in the past.

I haven't seen any issues with the leaves, do you have a screenshot or a short clip?

Metro Exodus EE runs faster and looks better than the original release. If there is room for similar optimization, then it should be done. The idea is to code taking advantage fully of the current hardware and new opportunities.

The engine/game (Indiana Jones, I mean), isn't quite perfect. The devs said with DF that the water for instance, has part of the code for PT reflections, but with everything going on in the jungle, performance isn't there yet ( ergo, they're not enabled), so it needs further polish that they hope will come (relative) soon. On the same train of thought, perhaps other stuff like latency can be improved, too.

Not the least, plenty of game that don't require the thick jungle environment, so performance would be significantly higher as it is in IG as well - with 4k Qualty around 60fps +/- being a possibilty even on 4080.

 

[Calin Banc]

Even a 4070 does decently.

Watching a capped 60fps YT vid doesn't=decent gameplay.

66-78 fps DLSS3, I'd NEVER run that fps on FG, using your example, my 4070 can technically do PT.

Drop to 1080p, DLSS Balanced and will be decently.

Don't forget people are playing on consoles in worse conditions, so if tens of millions are fine with less, surely this is "decent".

 

[Rroff]

There's no complex vegetation in Quake 2, though? As in, there's no vegetation at all, it's 99% indoor game.


And yet we've seen no sight of any of that in other games with PT either? Unless you count in RR, which has issues too and doesn't remove the problem fully either (even best AI won't help with just too low resolution and too much noise). MegaLights dev mentioned one has to be very careful designing games with RT to not overwhelm denoiser as is and that's not even full PT either. One would hope new AI based stuff will help but it cost performance too, so - as usual - it will have to be a balance between performance and quality.


A lot higher ray count means what? 4 instead of 2? We currently see about 4-5 max ray counts per pixel in best current RT and PT implementations. We need orders of magnitude more to have proper quality without AI dreaming up missing things. Sure, devs could actually do their job and optimise stuff much better than they currently do (on average) and maybe then we could get a few more rays or xx60 class GPU could handle it ok with current quality. But no, GPUs can't handle higher counts at all (not even 4090) - I have described earlier what AMD said is needed to combat that and it's not about transistors in the GPU anymore (or GPU performance in general). It's about memory speeds apparently. Especially if you want to connect it with AI processing - GDDR 7 is orders of magnitude too slow to handle that, so is L1 cache in itself apparently. New clever design of GPUs is needed to improve things.

Nothing complex in the vanilla game, but I've been doing some modding including testing vegetation, etc.

Some games have started to implement Ray Reconstruction but it is early days yet, same for stuff like RESTIR but there is a lot of working going on behind the scenes which will eventually make it into games - some guy has a fork of Quake 2 RTX which reduces noise considerably. Increasing ray count all helps even a relatively small increase can make a big difference with the right algorithms in place.

EDIT: The biggest cheat with Quake 2 RTX is that only the sun light does multiple bounces with full colour transfer, etc. all other lights are running much more constrained.

 

[WARburton]

when i see this two guys i just think left.... BELL, and right mega END.

after watching them smudge bench mars for years to advance intel no matter what they can sod off

 

[tommybhoy]

Drop to 1080p, DLSS Balanced and will be decently.

Don't forget people are playing on consoles in worse conditions, so if tens of millions are fine with less, surely this is "decent".

The easy reply was 'I didn't notice it was running FG'.


I didn't spend upwards of 2 and a half grand for running games upscaled from 50pence to 1080p to run settings that reduces the enjoyment of that investment.

I'll save everyone's time by not wasting my time uploading pics with Cyberpunk running full screen QHDUW on suitable settings with zero RT'ing Vs a 1080p windowed CP running PT'ing on a QHDUW panel.

If I want console gaming, I'll push the button on my PS controller that funnily enough would help mask that FG latency with the other 100 of millions of users that paid less than the cost of a 4070 to just play games.:thumbsup

 

[D3K]

when i see this two guys i just think left.... BELL, and right mega END.

after watching them smudge bench mars for years to advance intel no matter what they can sod off

Chuckle brothers. Their first video posted in here I managed a couple of minutes past their admission that they're biased against RT. They went on to show how, despite them being a hardware channel, that they had no idea how RT worked before 2000 series, lamenting the marketing of nvidia fooling people. Their last video on noise I didn't even bother as trust in their opinion is at rock bottom, and going by the chatter in here from some it's as if HUB somehow revealed a GPU vendor conspiracy, despite constant and well advertised updates to drivers to correct it wherever they can.
These two are not people to refer to on the topic of RT. Stick with DF support opinions with facts over feelings.

 

[Bidley]

when i see this two guys i just think left.... BELL, and right mega END.

after watching them smudge bench mars for years to advance intel no matter what they can sod off

HUB have been accused of favouring AMD, Nvidia and now Intel on this very forum in the past couple of weeks. Which probably tells you they're probably about right.

With regards to RT, it's a fact that it tanks performance, and there is visual noise that RR both fixes and introduces. In fact (if you watch any of their content) they do praise RT fairly often (Tim more so than Steve).

I really wish there was a little more room for nuance when talking about this stuff - RT in Cyberpunk is IMO an absolute masterpiece at times, the reflections look amazing, but the visual noise can be distracting, RR makes it worse for me, and sometimes raster shadows plain look better than RT shadows. And this is one game that has a (largely) good implementation, you can count the rest on one hand.

It remains to be seen how RT progresses in terms of uptake and performance/quality improvement. But it's been 6 years so it's slow going however you look at it.

 

[Calin Banc]

The easy reply was 'I didn't notice it was running FG'.


I didn't spend upwards of 2 and a half grand for running games upscaled from 50pence to 1080p to run settings that reduces the enjoyment of that investment.

I'll save everyone's time by not wasting my time uploading pics with Cyberpunk running full screen QHDUW on suitable settings with zero RT'ing Vs a 1080p windowed CP running PT'ing on a QHDUW panel.

If I want console gaming, I'll push the button on my PS controller that funnily enough would help mask that FG latency with the other 100 of millions of users that paid less than the cost of a 4070 to just play games.:thumbsup

I noticed. That 4070 isn't for 1440p, is for 1080p. This isn't your coup of tea which is fine.

 

[WARburton]

I noticed. That 4070 isn't for 1440p, is for 1080p. This isn't your coup of tea which is fine.

i had a 4070 about 14 months ago it was better at 1440 that 1080, but not good for RT at ever res.

 

[WARburton]

Drop to 1080p, DLSS Balanced and will be decently.

1080p DLSS Balanced would be 720 right? or is it lower