Discussion Nvidia Blackwell in Q1-2025

[MrMPFR]

I hope we get one or two new things based on that:
- DLSS automatically disables ingame denoisers (e.g. based on whitelisting and using existing game engine console commands). At least for UE5 this would be already quite helpful.
- We get at least a tool with a switch to disable ingame denoisers

Yeah they need to add automatic override to NVIDIA App when selecting Preset L and M, but I'm not sure that's possible and it seems like this new RR model is very inconsistent and breaks in many instances.

More inconsistent than SR branch. DLSS 4.5 just isn't a serious release and should be ignored.

 

[basix]

As I see it, denoising is a natural consequence of temporal supersampling. It just gets especially visible with stochastic rendering methods like RT which are inherently noisy (e.g. Monte Carlo sampling).
The better you can track information (pixel or even sub-pixel accuracy), the less noise will be visible in the final image. This can easily be tested by standing still in games vs. moving.

Here is a very neat comparison of DLSS 4.5 versus FSR 3.1. Both with enabled & disabled input denoiser:

FSR 3.1 looks pretty good when standing still, but when moving noise increases because information tracking is not optimal. If you get blurry input instead and therefore also blurry output you do not notice that to that degree.
Now, DLSS 4.5 is much better at information tracking. Therefore, the picture looks better than FSR 3.1 and much better when moving. If the denoiser is enabled, the difference is smaller while moving.

The problem with input denosing is, that denoising at lower rendering resolutions gives the denoiser algorithm a hard time. The results are noise, boiling artifacts and/or blur. You can trade one of those for eachother.
DLSS will have problems to resolve that, because it might interpret it as changing imagery and not noise. Or information might simply not be there anymore (blurred out).

Now the fun part:
DLSS 4.5 might be that good in tracking information, that some denoiser induced artifacts do not get suppressed or blurred anymore. This sounds counter intuitive, I know.
DLSS 4.0 was maybe just blurring that stuff and therefore it was not as noticable as now with DLSS 4.5.
This is just a hypothesis but it somehow would make sense.

Or in other words:
DLSS 4.5 is so good at information tracking, that the input denoiser is not required anymore. This is an exaggerated statement, there are still failing cases. But that seems to be the direction with better DLSS and FSR algorithms in the future.
I could imagine that DLSS 4.5 has a cleanup / lightweight denoise post-pass at full resolution. But this would inevitably add some blur to the image.

Here another cool use case:
Use DLSS preset L as denoiser for Dragons Dogma 2 with pathtracing.
Also DLSS 4 preset K denoises some parts of the picture (e.g. the floor) but fails with more difficult places like areas with only few light.

 

[MrMPFR]

Use DLSS preset L as denoiser for Dragons Dogma 2 with pathtracing.

Indeed. Here's the comparison with preset K:

Sizeable improvement.

 

[Heartbreaker]

As I see it, denoising is a natural consequence of temporal supersampling. It just gets especially visible with stochastic rendering methods like RT which are inherently noisy (e.g. Monte Carlo sampling).
The better you can track information (pixel or even sub-pixel accuracy), the less noise will be visible in the final image. This can easily be tested by standing still in games vs. moving.

Here is a very neat comparison of DLSS 4.5 versus FSR 3.1. Both with enabled & disabled input denoiser:

FSR 3.1 looks pretty good when standing still, but when moving noise increases because information tracking is not optimal. If you get blurry input instead and therefore also blurry output you do not notice that to that degree.
Now, DLSS 4.5 is much better at information tracking. Therefore, the picture looks better than FSR 3.1 and much better when moving. If the denoiser is enabled, the difference is smaller while moving.

The problem with input denosing is, that denoising at lower rendering resolutions gives the denoiser algorithm a hard time. The results are noise, boiling artifacts and/or blur. You can trade one of those for eachother.
DLSS will have problems to resolve that, because it might interpret it as changing imagery and not noise. Or information might simply not be there anymore (blurred out).

Now the fun part:
DLSS 4.5 might be that good in tracking information, that some denoiser induced artifacts do not get suppressed or blurred anymore. This sounds counter intuitive, I know.
DLSS 4.0 was maybe just blurring that stuff and therefore it was not as noticable as now with DLSS 4.5.
This is just a hypothesis but it somehow would make sense.

Or in other words:
DLSS 4.5 is so good at information tracking, that the input denoiser is not required anymore. This is an exaggerated statement, there are still failing cases. But that seems to be the direction with better DLSS and FSR algorithms in the future.
I could imagine that DLSS 4.5 has a cleanup / lightweight denoise post-pass at full resolution. But this would inevitably add some blur to the image.

Here another cool use case:
Use DLSS preset L as denoiser for Dragons Dogma 2 with pathtracing.
Also DLSS 4 preset K denoises some parts of the picture (e.g. the floor) but fails with more difficult places like areas with only few light.

There is some good work in the new presets, but it seems like it really needs developers to specifically tweak for it. So I'm just going to ignore 4.5. If it's worth it hopefully the Dev includes it.

 

[MrMPFR]

@basix a shame it doesn't hold up in RTX Remix at all:

Same issue in HL2 RTX I've been told.
But does resolve non-PT RT reflections remarkably well, well at least in UE5 games.

There is some good work in the new presets, but it seems like it really needs developers to specifically tweak for it. So I'm just going to ignore 4.5. If it's worth it hopefully the Dev includes it.

I agree visuals with RT being broken out of the box is just a joke. Shouldn't be on the consumer to fix all these problems + they should've been more frank about the issues with RT lighting or delayed the model until it was fixed.
Maybe a slim chance of an April SR fix when DGF launches.

Also based on just how poor perf is on 20-30 series I doubt devs want to have anything to do with DLSS 4.5. Don't think we'll see any new games shipping with it.

 

[Heartbreaker]

I agree visuals with RT being broken out of the box is just a joke. Shouldn't be on the consumer to fix all these problems + they should've been more frank about the issues with RT lighting or delayed the model until it was fixed.

AFAIK, you won't get the new models by default. You have to use the NVidia app override to activate them. So you would have to first cause the problem to need to fix it.

 

[basix]

@basix a shame it doesn't hold up in RTX Remix at all:

Same issue in HL2 RTX I've been told.
But does resolve non-PT RT reflections remarkably well, well at least in UE5 games.

Is that DLSS 4 RR? It looks kinda blurry. Preset L shows so much more inner surface and texture detail as well much more detailed and crisp shadows.

 

[MrMPFR]

AFAIK, you won't get the new models by default. You have to use the NVidia app override to activate them. So you would have to first cause the problem to need to fix it.

Indeed but by using override function in NVIDIA App something they've pushed for +1 year you do.

Is that DLSS 4 RR? It looks kinda blurry. Preset L shows so much more inner surface and texture detail as well much more detailed and crisp shadows.

Yeah DLSS RR does that. IIRC Hardware Unboxed said it has problems with texture detail and is blurs in motion.
Here is an example where Preset L struggles:

https://www.reddit.com/r/nvidia/comments/1qxwtp0/dlss_45s_denoising_ability_isnt_quite_a_full_ray

 

[basix]

In that video it looks much less severe for RR.

But look here at the shadow detail, the texture detail and speculars. So much missing from the RR version. You are sure that this is DLSS 4 RR and not DLSS 3 RR (even if they say DLSS RR Transformer there)?

Spoiler: RR vs. Preset L

If this is DLSS 4 Transformer, we have still a long way to go. Maybe denoising at output resolution and not input resolution?

 

[MrMPFR]

But look here at the shadow detail, the texture detail and speculars. So much missing from the RR version. You are sure that this is DLSS 4 RR and not DLSS 3 RR (even if they say DLSS RR Transformer there)?

It says TF RR, but can't be 100% certain. But I've heard people, including HUB, say DLSS 4 RR has issues with texture detail.

Maybe denoising at output resolution and not input resolution?

No idea.

Hopefully DLSS5 can fix ALL the issues: SR, RR and FG-related.

 

[basix]

It says TF RR, but can't be 100% certain. But I've heard people, including HUB, say DLSS 4 RR has issues with texture detail.

I tested by myself today with Portal RTX: RR with preset E, SR with preset M. The override to L somehow did not work, L was only applied in Ultra Performance mode and K ind balanced/quality mode. Other games did rightfully appliy preset L everywhere.

It really depends on the angle of the incident light. Depending on that, there might be is a stark difference in details. RR is clean but lacks details.
Preset M was able to resolve more shadow, texture and particle details. RR was much better at reflections and ghosting suppression. And some dim and changing light sources are visible with M/L but not at all with RR. RR is good but does still attenuate or even remove many details.
All this stays true with preset M/L + standard denoiser. Texture and some shadow / specular details are much better than RR. RR keeps the lead with some reflections, but not reflections with inner texture details.

My preferred way to play Portal RTX would now be preset M/L + standard denoiser. RR has some benefits but you lose details.

Very interesting to see that now compared to preset M and L, as RR Transformer got somehow considered as gold standard of temporal upsampling. It is very good but not yet perfect.

Hopefully DLSS5 can fix ALL the issues: SR, RR and FG-related.

I hope so, yes.

I hope the same for FSR 5 and XeSS 4

 

[basix]

I made a comparison of RR (Preset E) / Preset M / Preset M + Denoiser (4K DLSS Performance) in Portal RTX:

1. https://imgsli.com/NDQ4MjQy
2. https://imgsli.com/NDQ4MjQz
3. https://imgsli.com/NDQ4MjQ0

My observations:

• RR shows in general more accurate lighting, reflections and shadow details (with one exception, more on that later)
• Preset M is able to extract much more fine detail, especially on flat surfaces
• If you compare RR with preset M without denoiser, RR shadow and caustic & reflection details are very close to that non-denoised "ground truth"
  • RR seems to be able to preserve lighting detail very well
• The big drawback of RR is, that inner surface detail is missing. DLSS 4.5 shows much more texture and specular detail (even including denoising)
  • Also DLSS 4.0 preset K + Denoiser has an edge over RR in this regard
  • RR seems to trade visibility of high frequency details with stability and cleanlyness
• RR preset E still uses logarithmic color space. This is especially visible with the bright energy ball, where the coloring of the scene is quite different (white-ish with RR, orange-ish with preset M)
• Even if RR dynamic shadows are good, it is still quite blurry and diffuse and lacks crispness vs. non-denoised preset M
  • Not sure if this is because of general RR blur or because logarithmic vs. linear color space. If the bright energy ball brightness is compressed to a much dimmer value, the shadows would be much softer as well.
• Preset M does better with bright particles (again: logarithmic vs. linear color space)
• The standard denoiser does completely mess up with some reflections and shadows. They are simply wrong or missing

I think a DLSS 4.5 version of RR could already get rid of the biggest flaws of DLSS 4.0 Preset E. Together with an adjusted denoising scheme to allow more inner surface detail (less aggressive denoising, but that should be possible with the enhanced DLSS 4.5 models).
I hope Nvidia releases such a model in the upcoming weeks, maybe together with 6x MFG.

 

[MrMPFR]

Thank you for the insights. The stuff about surface detail is a well known issue that Hardware Unboxed talked about a year ago. Blurriness in motion is also an issue, as bad as CNN DLSS.

I hope Nvidia releases such a model in the upcoming weeks, maybe together with 6x MFG.

I'm worried about that this is impossible without NVFP4 and stronger HW. Hopefully I'm wrong.

 

[basix]

I would not mind if Nvidia brings it with NVFP4 (and FP8 / FP16 fallback on older architectures).
Sure, it will cost you framerate, but you always have to option to go back to preset E. I rather want to have the option and decide to not use it (because of...) instead of not having the option at all

But anyways, I do not know what the difference between RR and rrlite / rrlite_folded actually are. Maybe there is only a minor difference, we simply do not know.

One additional note regarding performance:
Preset M + standard denoiser results in the same framerate as plain RR.

 

[dangerman1337]

https://videocardz.com/newz/rtx-titan-blackwell-or-rtx-5090-ti-nvidia-reportedly-is-developing-a-new-rtx-50-series-halo-gpu-for-q3-2026

Sounds probably the same yielding die as RTX 6000 Blackwell Pro but certainly way less VRAM, probably 32 or 48GB depending what GDDR7 modules they use and they're probably positioning this as something to go with top Zen 6 or Nova Lake-S SKUs (like the planned bLLC NVL-S ones). Makes sense to me because if you can afford a NVL-S bLLC CPU+RAM+top 900 series motherboard you could afford a 5090 Ti of sorts.

Though I wish they did a 5090 and 5090 Ti in the first place :/, hopefully a sign they'll just do that day 1 with Rubin with a 6090 Ti and 6090 Vs RDNA 5's AT0 (Jensen doesn't wanna lose). Cleaner lineup (Ti full-ish die, non-Ti cut down) please in the future than the garbled messes we've gotten.