Digital Foundry Article Technical Discussion [2024]

[DavidGraham]

Can anybody take Nvidia's claim that they're actually doing "path tracing" at that point with their RTX branch of UE5 if there's major inaccuracies like their need to do screen space hacks ?

So you don't like a possible partial minuscule screen space shadow here and there in intersections, and consider this a flaw and a hack, while at the same time you ignore the orgy of screen space effects flashing before your eyes with the stock UE5 solution on everything from entire reflections to entire shadows to ambient occlusion to global illumination? Worse yet you consider this orgy of hacks the superior solution?

 

[pjbliverpool]

How much gflops was G80 and how much Cell+RSX? ~400 and ~400. Of course G80 was more advanced and easier to be programmed.

Even back in those days raw FLOPS was a meaningless number. Depending on how you count it the 8800GTX was rated at either 345GFLOPS or 513GFLOPS (it could do a 3rd shader operation per clock under specific circumstances). Cell + RSX would have been 436GFLOPS and the X1900XTX would have been 426GFLOPS. So both of the PC GPU's are arguably lower in that metric overall than the PS3 but in literally any game they would have both obliterated it in real world performance.

Yes, but in long terms no. X1900XTX wasn't able to run DX11 games.

Neither was the Xbox 360? At least not with the full DX11 (or DX10) feature sets. Xenos did go further towards the full DX10 spec than the X1900XTX (including unified shaders) but that was it's single advantage over the PC GPU IMO. And in the real world it didn't budge the needle as to which was the more effective GPU.

I think the first DX10 only game didn't release until over 5 years after the 360 and XTX launched and you're only talking a very small hand full of games in that category (including DX11) during the 360's entire lifespan. How does that compare to the many thousands of games that the XTX would have played better over the same period?

 

[Lurkmass]

So you don't like a possible partial minuscule screen space shadow here and there in intersections, and consider this a flaw and a hack, while at the same time you ignore the orgy of screen space effects flashing before your eyes with the stock UE5 solution on everything from entire reflections to entire shadows to ambient occlusion to global illumination? Worse yet you consider this orgy of hacks the superior solution?

Possible ?! There was no room for ambiguity where NV developers themselves have explicitly stated beyond any doubt that they *absolutely* use screen space techniques as fallbacks highly geometrically dense Nanite meshes which might explain the several observations seen with the weak GI results in the game ...

I trust Andrew's judgement w/ respect to his statement (since he's seen the very code himself) when he professes that the RTX branch doesn't "diverge" as Nvidia wants people out there to believe and especially when the technical press gives them a pass for their marketing machine on far too many occasions wo/ thorough critical reviews. Who else knows what other screen space methods they rely on in their so called "path tracing" implementation besides just shadows alone ?

 

[DavidGraham]

Possible ?! There was no room for ambiguity where NV developers themselves have explicitly stated beyond any doubt that they *absolutely* use screen space techniques as fallbacks highly geometrically dense Nanite meshes which might explain the several observations seen with the weak GI results in the game ...

Firstly, this was specifically in relation to RTXDI, not any thing else, and it was stated it "could" result in a less accurate shadows in intersections, it's a possibility that varies depending on the geometric complexity of the model. So this only applies to specific shadows. Not all shadows not all objects, but specific shadows on specific parts of specific models.

Secondly, the path traced illumination solution looks leaps and bounds better than the standard Lumen solution. So I don't get your comment about weak GI, this should be directed at Lumen.

Who else knows what other screen space methods they rely on in their so called "path tracing" implementation besides just shadows alone ?

Of course you will trust whatever will suit your guesswork theories/narratives, but that doesn't change the fact that you consider the party of screen space effects (Lumen) better than path tracing, which is a lopsided logic, and doesn't change the fact that you used incorrect statements regarding the use of Nanite in the game.

that they don't subscribe to it's full power ...

That's not news, it's 99% of UE5 games.

 

[trinibwoy]

Possible ?! There was no room for ambiguity where NV developers themselves have explicitly stated beyond any doubt that they *absolutely* use screen space techniques as fallbacks highly geometrically dense Nanite meshes which might explain the several observations seen with the weak GI results in the game ...

I trust Andrew's judgement w/ respect to his statement (since he's seen the very code himself) when he professes that the RTX branch doesn't "diverge" as Nvidia wants people out there to believe and especially when the technical press gives them a pass for their marketing machine on far too many occasions wo/ thorough critical reviews. Who else knows what other screen space methods they rely on in their so called "path tracing" implementation besides just shadows alone ?

Yes they’re likely using screen space techniques to fill holes. Not sure what’s so damming about that or how it diminishes the benefits of PT. Are you suggesting that path tracing implementations must be perfect in order to show progress? 3D rendering is full of hacks and always has been.

 

[DavidGraham]

Very good analysis of the PC tech in Star Wars Outlaws.

 

[Jupiter]

I play it and find the lighting very impressive thanks to GI and RTXDI. I'll have to think about what comes after Cyberpunk 2077 Pathtracing in terms of lighting. Alan Wake 2 or Star Wars Outlaws.

Of the new games I've played this year, Star Wars Outlaws is probably number 1 in terms of graphics. The fact that all the light sources now cast shadows is technically a big improvement over Avatar. I find it incredibly annoying in most games when many shadows from local light sources are just mmissing. Hard shadows also simply look very outdated and gamey.

The sound is also very good in Star Wars Outlaws. Avatar had the best ambient sound out of all the games I've played. It sounds like a new kind of sound. Less video game-like and more realistic, physical, but not boring either. Cinematically realistic I would call it. I also think Battlefield 2042 is exceptional when it comes to mixing the sounds of war. In Battlefield 2042 is a good mix and selection and no wild confusion like in other games where many players fight each other. That's my favourite time to turn up the volume of my enthusiast sound system.

The difference between console and PC was also mentioned in the thread. So it's clear to me that the difference between PlayStation 3 and PC was enormous back then. With a GT 8800 you were already in a different generation when it comes to graphics and effects. The difference between PC and PlayStation 4 was much smaller. In the PlayStation 4 era most games on the consoles did not lack countless graphical effects. Tessellation was one of the few things that wasn't so good there.

 

[Charlietus]

The difference between console and PC was also mentioned in the thread. So it's clear to me that the difference between PlayStation 3 and PC was enormous back then. With a GT 8800 you were already in a different generation when it comes to graphics and effects. The difference between PC and PlayStation 4 was much smaller. In the PlayStation 4 era most games on the consoles did not lack countless graphical effects. Tessellation was one of the few things that wasn't so good there.

I wonder what Sony and Microsoft were thinking before launching this generation. They knew what was possible on those machines, and the small upgrade to graphics they would offer compared to past generations. The specs are fine, but the technological ambition was small to inexistent.

 

[Andrew Lauritzen]

Sigh... I really hesitate to wade back into this but given people are directly attributing quotes to me I will at least try and clarify those bits.

The game doesn't even make use of VSM as well because testing the clipmaps against large primitives would involve high overhead due to reduced culling efficiency ...

I would have assumed this as well, with non-Nanite terrain being a big issue usually as well. That said, you can enable VSMs in the ini files and it is only very slightly slower than the shadow map path and so far I haven't seen any artifacts that aren't related to some foliage LOD/shadow proxy thing they have going on and thus are the same in the regular SM path as well. I really don't know why they didn't provide them as an option, as they bring the visuals a lot closer to the RT path with a much smaller performance hit.

I will wait the final judgment on that, that very same RenderDoc gave a false result for Lumen.

Ok, the initial test was not invalid because of RenderDoc or the person who did it - it's because the game is buggy with settings changes sometimes so when he took the capture it literally was running without Lumen (i.e. GI "low"). This is interesting to note because it now seems highly probably related to...

As for Path Tracing and Nanite, I think we can all agree that the results shown so far are anything but bad.

Yeah so... there are spots where it looks good, but the irony is the ones you kept pointing out in other threads from online videos are not them. In fact, given the above I've now connected some dots in my head and I'm actually pretty certain that those examples where people keep pointing out "oh it's so much brighter with RT" (but it looks totally flat) are actually bugged settings - probably the same bug the original RenderDoc capture hit - and you could probably get the same result by forcing off GI (with either RT setting). I have tried to be polite, but honestly if you look at those images and think that they are great looking lighting/PT then I question your ability to judge what looks good in the first place. Nothing smells brand loyal/fanboy nonsense more than literally looking at a constant ambient term but promoting the difference as the best thing in the world because someone told you it was path tracing.

Yes they’re likely using screen space techniques to fill holes. Not sure what’s so damming about that or how it diminishes the benefits of PT. Are you suggesting that path tracing implementations must be perfect in order to show progress? 3D rendering is full of hacks and always has been.

Yes of course they use screen space traces to deal with the lower res RT meshes, for all types of raytracing. This is completely normal and expected and not something to get upset over. That said, it's just as silly to get upset over screen space traces in Lumen. Particularly since the screen space tracing code they use in the RTX branch is literally lifted from/exactly the same as the Lumen trace code.

Obviously triangle RT is generally higher quality than SDF RT, but Lumen can of course do both. Implying that Lumen can only do screen space reflections and/or SDFs is silliness - that's a performance tradeoff that games decide. Which brings to the next point...

Andrew raised some concerns about GI quality. Haven’t seen any complaints about the shadow quality. He mentioned RT shadows were static but that doesn’t seem accurate based on released footage.

See above for "GI quality". Shadows look totally fine in the final game to me. The caustics are neat and I like them. The main concession is of course...

The performance is reasonable and is mostly inline with other path traced titles that don't rely on virtualized geometry.

"In line" or not, it really needs to be emphasized that it's way slower. Like really barely usable on a 4090 really fucking slow. Like even me with my 4090 who loves pretty graphics and hates bad shadows will not play it at those frame rates. I don't know why we keep circling around on these same things to be honest... the more approximate techniques are faster, the more accurate ones are slower. It's a good thing to have options all across the curve. It's probably still going to be a long time before a AAA game could ship with only a RT path the cost of Wukong's, especially considering consoles. It's great that I can maybe play with RT on a 5090 or whatever in the future, but let's not pretend that the two paths are comparable at all in terms of performance.

They locked in on 5.0 to get the game out as soon as possible. I get it, the game has a massive amount of content that would need to be retested-redone.

I'm still skeptical that this is actually 5.0. Is this based on anything other than binary metadata yet? Someone could probably get a better estimate by trying some cvars that were added in 5.1, 5.2, etc. and seeing which are present. I think some of the Unreal tools that enable the console or similar could probably dump the full list for cross referencing. Or you know, someone could ask the developers

 

[DavidGraham]

"In line" or not, it really needs to be emphasized that it's way slower.

Here lies the difference, the argument was that Nanite + Ray Tracing was so prohibitively expensive that it couldn't be done with acceptable performance, that argument is proven wrong by the actual integration in Wukong, not only did the developer implement ray tracing, but they implemented many ray tracing effects (illumination, shadows, reflections, caustics). And with good results.

Like really barely usable on a 4090 really fucking slow

It's not signifcantly slower than other non nanite path traced games (Portal RTX, Cyberpunk, Alan Wake 2, Call of Duty Modern Warfare 3, etc). That's the whole point.

I have tried to be polite, but honestly if you look at those images and think that they are great looking lighting/PT then I question your ability to judge what looks good in the first place

Let's remain polite, and agree to disagree, you have said before that other areas show a clear difference, the areas you and I disagree about remain debatable, I see a clear lack of bounce lighting in several obvious spots of the image, this couldn't result from a change in Ambient Term alone.

So unless you have a clear evidence that it is the Ambience difference I hope we leave it at that.

 

[Andrew Lauritzen]

Here lies the difference, the argument was that Nanite + Ray Tracing was so prohibitively expensive that it couldn't be done with acceptable performance, that argument is proven wrong by the actual integration in Wukong, not only did the developer implement ray tracing, but they implemented many ray tracing effects (illumination, shadows, reflections, caustics). And with good results.

Right because it's not doing "RT + nanite" in any real way. It's doing RT on a low res approximation of the scene, and you can tell in the lighting if you know where to look. You could try "r.RayTracing.Nanite.Mode 1", but I would not be surprised if it crashes or is even more slow due to the overhead. Ultimately we really want it to be doing RT on the same quality geometry that is being rasterized in screen space, especially for shadows and mirror reflections. For now though using fallback meshes is a reasonable tradeoff until we have better control over acceleration structure builds and streaming. For GI specifically the importance of fine-scale detail is lower which is incidentally why even SDFs do okay in a lot of cases.

Let's remain polite, and agree to disagree, you have said before that other areas show a clear difference, the areas you and I disagree about remain debatable, I see a clear lack of bounce lighting in several obvious spots of the image, this couldn't result from a change in Ambient Term alone.

I'm talking specifically about the so-called "path tracing" result in those video comparisons you linked. It is clearly not doing path tracing - and probably not any GI; something is busted. If you think it looks like path tracing, then I - politely - I don't trust your opinion on GI.

 

[Shifty Geezer]

I haven't followed this closely but of the images I've seen, you have one version with opaque shadows that aren't even black and the other with detail in the shadows that isn't in keeping with correct optical contrast.

I think the conversation would be helped by someone taking a few A/B images and expressing their interpretation, and opening it up to the floor to see if there's a consensus on what we're seeing.

 

[DavidGraham]

It's doing RT on a low res approximation of the scene

Yes it is, the argument was even that is expensive and can not be done, which is obviously incorrect.

I'm talking specifically about the so-called "path tracing" result. It is clearly not doing path tracing

It think -as you have stated- it's quite easy to prove your theory right and my interpretation wrong, just turn off Lumen in both scenarios and compare the results.

I would be happy if I, Digital Foundry, other tech publications, youtubers and NVIDIA are proven wrong if the game is relying on ambient term to do it's global illumination.

I think the conversation would be helped by someone taking a few A/B images and expressing their interpretation, and opening it up to the floor to see if there's a consensus on what we're seeing.

I would like that. I did that a few posts back, I also posted videos comparing the two, but I am happy to provide more.

 

[Andrew Lauritzen]

It think -as you have stated- it's quite easy to prove your theory right and my interpretation wrong, just turn off Lumen in both scenarios and compare the results.

Wait, are you only looking at this game via videos? I doubt it's 100% reliable to trigger the bug... it's clearly not supposed to be doing what it is.

In any case I think *specifically those video comparisons* are busted, probably due to settings change bugs. Same situation with the original RenderDoc capture where both Lumen and RT were off. I don't think the general implementation is always broken - it has seemed fine the times I've tested it, but not exactly night and day differences (pun intended this time).

I guess the discussion was in the other thread, but I'll repost the one image as it's the most stark:
I assume we can all agree that this is clearly wrong/no occlusion skylight/no GI or something similar, yes? That whole video seemed busted. The PC-Cine results looked plausibly busted in similar ways although it's possible the "full RT" was flat due to being low res in that instead since the indirectly lit regions were smaller in the shots they took. I'd probably put my money on at least some of their comparisons being broken too though... there was nothing that screamed great indirect lighting in any of those comparisons either.

The main image that looks clearly better that I've seen posted is the one from the NVIDIA blog, and I confirmed that one locally as well.

That location clearly has better indirect specular among other things in the "full RT" result. But notably there's no massive difference in total radiance, nor is there entirely missing indirect occlusion in the RT result.

 

[Shifty Geezer]

Full images for convenience...





Why's the exterior completely different? Way more contrast on the RTX Off version. the mossy stone just outside is the same intensity for both.

Otherwise I agree with Andrew on this point, total luminance in that architecture looks the same. Also in another RTX screenshots, everything except one feature is identical...





Visual delta:

Away from the reflections there's very little difference.

I think some decent RTX On/Off of more select points would be beneficial. Some of the most egregious points in the video are maybe the fault of the capture process? So much washed out darks and black crush! It's like PS360 struggling to sort out HDTV again.

 

[techuse]

I believe Andrew was referring to this comparison as suffering from the settings bug. I have time stamped it.

 

[DavidGraham]

I assume we can all agree that this is clearly wrong/no occlusion skylight/no GI or something similar, yes?

The RT solution in the game is a 2 bounce solution, it's possible that statue is too deep inside the cave to be affected by the 2 bounce solution, it could need more bounces. It's possible.

However, here are more comparisons of the stark difference. Left is RT on, right is off, open in new tap for a bigger picture or just click on the pictures and press the arrow keys to quickly move through them and compare.

1st comparison: more light coverage from the fire on the stones and tress, more bounce light over the character and the stones to his left.

2nd comparison: more bounce light on the character and every rock, tree, grass on the ground to his left and on the mountain.

3rd comparison: more bounce light on the building in the distance, and more depth on the tree in the center plus more color bounce on the foliage, the RT off picture darkens the building in the distance and have no depth for the trees or color bounce.

I think it can't be more clear than this. There are more comparisons in this link, but I chose the ones focusing on illumination rather than shadows or reflections.

 

[Davros]

Richard Leadbetter from digital foundry yesterday

 

[trinibwoy]

Very good analysis of the PC tech in Star Wars Outlaws.

So there’s no RT shadows aside from RTXDI? Then this blurb from Nvidia doesn’t make sense.

As part of our collaboration, Star Wars™ Outlaws on PC is also being enhanced with NVIDIA RTX Dynamic Illumination (RTXDI), which complements the game’s ray-traced reflections, ray-traced shadows, and ray-traced global illumination, taking visuals to the next level.

I hope RTXDI catches on. The lack of shadows from most lights is a glaring issue in modern gaming. We just got used to it. Time to move on. The DF video highlighted another benefit that isn’t discussed as much. A bunch more lights actually cast light too especially area lights which aren’t raster friendly.

 

[Andrew Lauritzen]

The RT solution in the game is a 2 bounce solution, it's possible that statue is too deep inside the cave to be affected by the 2 bounce solution, it could need more bounces. It's possible.

Then it would be dark, not bright. The exposure on the direct lit stuff is the same in the videos that you can see - the RT solution is very clearly wrong/off in those comparisons.

However, here are more comparisons of the stark difference.

These are better comparisons that at least pass the "not clearly broken" glance, but still raise a few questions.

1st comparison: more light coverage from the fire on the stones and tress, more bounce light over the character and the stones to his left.

There's actually *less* indirect illumination overall in the RT image, it's just at a higher exposure. You can tell by comparing the direct illumination of the fire - which you seem to have interpreted as GI. It's not totally clear which one of these is more correct to the source data to be honest... weirdly the RT version is more contrasty and dark in the woods which is normally the opposite since SDFs tend to over-occlude with foliage.

2nd comparison: more bounce light on the character and every rock, tree, grass on the ground to his left and on the mountain.

The RT image is superior here. Note that unlike the videos you linked earlier there is actual indirect occlusion in the shadowed regions. You could probably still arrive at a broadly similar image with a mild tone map/exposure curve to be honest in this example, but I agree the RT image is better.

3rd comparison: more bounce light on the building in the distance, and more depth on the tree in the center plus more color bounce on the foliage, the RT off picture darkens the building in the distance and have no depth for the trees or color bounce.

The RT image is better, but has a few issues. First, it actually inverts the previous trend of indirect lighting where it has less in the shadowed part under the pavilion. I would not doubt that people would point that out as an advantage if it were on the RT side The tree shading is actually probably "wrong" on the RT side in that it's not handling transmission/subsurface. The raster approximation of these is very simple, but the content is clearly calling for it if you compare those leaves to the broadleaves. RT seems to be treating both as fully opaque. Not a big deal and not something you'd really notice unless you A/B side by side though - you'd just assume they were a different kind of leaf.

The practical question though is is it worth the performance hit for what I think many people would consider fairly minor differences. The RT result is better, but I don't know that it's worth halving performance for unless you're already well over 120fps. We're gonna need new GPUs for that unfortunately. I really do hope they provide separate options for RT shadows and reflections as that gets you the majority of the benefit in this game with presumably a fraction of the overhead.

I'd be curious how correct people would be if they had to blind A/B pick the path from the two images, particularly if you could have RT shadows (or even VSMs) in both as the CSMs are bit of a give-away.