• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

DLSS 5 preview

"You can just apply an orange filter" :cry: :cry: :cry: :cry: :cry: :cry: :cry:

Just went full circle, haha!

PS: I hope he's trolling, but probably he's not...
Well, it's likely a bit of both but he's right - there's very little visual difference in that specific example and huge huge FPS difference :) I can bet most people wouldn't be able to tell the difference when playing, which one is PT and which one is not, but FPS they do notice. And so, as he said, that's what dlss 5 could do - just put a filter on ;)
 
Last edited:
PS: I hope he's trolling, but probably he's not...

The first comparison he cuts off the bits at the side which show path tracing a bit better... not that path tracing is doing anything much in those scenes but that is more down to game design and implementation rather than the potential of the technology.
 
The first comparison he cuts off the bits at the side which show path tracing a bit better... not that path tracing is doing anything much in those scenes but that is more down to game design and implementation rather than the potential of the technology.
Nobody plays potential though; people play the implementation only. Same as DLSS 5 - in theory good potential but what they shown caused huge backlash. So, if it looks almost the same with 4x+ FPS difference, and also the PT version has big ghosting and noise issues (as he shown and we've all seen that in PT games), potential is meaningless IMHO.
 
Last edited:
He made a good point regarding motion vectors - it's not just for stability, but also if frames don't change much, DLSS 5 in theory does NOT have to generate those static parts for each frame and only need to do it for moving bits. This is likely why NVIDIA did NOT move much at all in presentations, most of the time they have shown near static scenes only. Likely their 2x5090 were still too slow to handle all of it with fully moving scenes, but if only small bits were moving it worked fine. This will likely be improved as time goes, but I still suspect it won't work great on current hardware below 5090 or 5080, with sensible resolution and FPS.

But that aside, most they talk about is some theoretical future DLSS 5+ version that will replace PT, generate geometry and do the whole neural rendering. That's not what currently shown DLSS 5 does, and not even the one CEO of NVIDIA described recently. That's some distant future for rtx6 or 7k+ series GPUs.

One thing to remind people - AI algorithms like DLSS 5 work fastest and most efficient on FP4 tensor cores and that's only in 5k series GPUs. 5k series are also the only ones where tensor cores have direct low-latency access to CUDA cores and so game can very quickly calculate things between tensor cores and CUDA cores forth and back with very low latency. 4k series and older do NOT have that. Which is why I highly doubt DLSS 5 will be any good/usable or even work at all on anything below 5k series.
 
Last edited:
It looks good, yet what about developers and artists and what they intended? Shame I just got 9070XT
it literally looks like a social media filter for games..... also what the artists intended is one of the big reasons people don't buy games.... you can already see an agenda underneath a lot of the time
 
Last edited:
Well, it's likely a bit of both but he's right - there's very little visual difference in that specific example and huge huge FPS difference :) I can bet most people wouldn't be able to tell the difference when playing, which one is PT and which one is not, but FPS they do notice. And so, as he said, that's what dlss 5 could do - just put a filter on ;)

Little visual difference, indeed. I bet the red/orange color cast on her hair can be had just with a similar color filter plaster over the image. Job done. And if you have multiple such colored light sources, you just mix and throw all those colors at the screen via filters. boom, done! :D

Grace-Resident-Evil-Raster-vs-RT-vs-PT.jpg



Even going by the "filter", from what we have now, a "smart filter" like DLSS5 needs 2x5090. Those two would run a whole lot faster with PT than such an "instagram filter".
 
Last edited:
My concern is that we'll end up with another rasterization lookalike, meaning just approximation of how light works by guessing based on the data it was trained instead of actually doing the path tracing that should be done normally...
That's a given but also that's why Nvidia engineer said clearly, it works best in pt games as it needs very good input of proper lighting to give good enough output. Ergo, that's going to make games even slower.
 
That's a given but also that's why Nvidia engineer said clearly, it works best in pt games as it needs very good input of proper lighting to give good enough output. Ergo, that's going to make games even slower.

It works best with PT/RT, because it's a screen-space solution, but with PT you already get the benefits of a better lighting that DLSS supposedly brings for the landscapes - even more so with mega geometry and opacity maps. It remains viable for neural materials (basically enhanced objects) and perhaps characters, but even there, a properly made game with future engines (to no have it limited just at UE5/6), will probably come close - if not surpass it.

It could help "lazy devs", but don't know, can't say I'm impressed. It would have been nice to help setup quickly a character, from a picture or a movie, into a game asset, a sort of mega humans from epic or have a sense of the world and create readily available assets for use in games - basically high quality generative AI for in-game assets (of course, running offline, not at run time)...


As of now, I find it interesting for less complex games (current and older), but pulled down with the limitations of screen-space. Can't say I'm incredibly impressed.
 
Last edited:
So
It works best with PT/RT, because it's a screen-space solution, but with PT you already get the benefits of a better lighting that DLSS supposedly brings for the landscapes - even more so with mega geometry and opacity maps. It remains viable for neural materials (basically enhanced objects) and perhaps characters, but even there, a properly made game with future engines (to no have it limited just at UE5/6), will probably come close - if not surpass it.

It could help "lazy devs", but don't know, can't say I'm impressed. It would have been nice to help setup quickly a character, from a picture or a movie, into a game asset, a sort of mega humans from epic or have a sense of the world and create readily available assets for use in games - basically high quality generative AI for in-game assets (of course, running offline, not at run time)...


As of now, I find it interesting for less complex games (current and older), but pulled down with the limitations of screen-space. Can't say I'm incredibly impressed.
In this case it seems we are aligned close enough in the opinion. I like how Jensen finally changed his tune and started to admit he doesn't like AI slop himself and can see how what was shown could be seen as that. But then he promised it will get much much better and give much more freedom to adjust by Devs etc. Sometime, in the future, who knows on what GPUs:)

The thing is that dlss5 in shown version is just a fancy filter, that's it, there's nothingn more to it. It's nothing like the neutral rendering they promised (textures compression, full 3D neutral faces, neutral geometry, neutral lightning replacing or speeding up considerably PT etc.). Those neutral things are nowhere to be seen after what they published in white papers and what we got instead is nothing like they described before. I'm still waiting for the real thing not the "sod off" glorified filter.
 
Last edited:
This guy (Ex dev from coffee stains studio) raised quite an interesting point. Will developers take DLSS5 into account and shift their entire artstyle to bring out the best from it, essentially alienating everyone who doesn't use.
Or will they tack it on at the end with no care to the output, which would annoy all of those who love their AI slop filter DLSS5 and how it looks.

The original DLSS was something you could just tack on the end, but DLSS5 could guide the whole art direction of the game from the very beginning. If it does actually have the effect of guiding the whole direction of the game then I doubly hope that DLSS5 fails.

 
Back
Top Bottom