NVIDIA DLSS 3.5 announced: Beyond denoising

[Middcore]

4 hours ago, porina said:

I think the only problem here is that we now have 3.0 requiring 40 series (or newer?) but 2.x and 3.5 do not. This could get confusing and maybe nvidia need to rebrand away from a version number scheme and explicitly call it by function.

 

Confusion about exactly how all of these whiz-bang proprietary features work and what you need to run them is, from Nvidia's perspective, a feature and not a bug, because it leads to people throwing up their hands and just buying the latest, most expensive card to be sure they'll have access to all of the latest features they can't possibly game without.

 

In 2019, when the Nvidia twitter account shared screenshots of the upcoming Red Dead Redemption 2 PC version and I saw a horde of fanboys in their replies spamming "#RTXon" despite the fact RDR2 didn't include a single "RTX" feature, I realized what the strategy was, and Nvidia has only doubled and tripled down on it since. 

[StDragon]

So if you have a 20 series, why not call it DLSS 2.5 since it can't do FG.

Confusing? Not nearly enough.

[Eigenvektor]

2 minutes ago, StDragon said:

So if you have a 20 series, why not call it DLSS 2.5 since it can't do FG.

Confusing? Not nearly enough.

But DLSS 3.5 can do Frame Generation. It simply requires the appropriate hardware support (i.e. a 40-series GPU), the same way DLSS 3 already required a 40-series GPU for that feature.

 

Without a 40-series GPU, it still supports Super Resolution and it still adds support for Ray Reconstruction.

[StDragon]

Just now, Eigenvektor said:

But DLSS 3.5 can do Frame Generation.

I should have clarified: Allow me to re-word that.

"So if you have a 20 series, why not call it DLSS 2.5 since the 20 series can't do FG."

[Avocado Diaboli]

I have to wonder why people find the versioning confusing. Let's look at an alternative to illustrate the point a bit clearer: Someone creates a digital audio workstation software. Version 1 can handle audio input from external devices that accept mic or instrument level audio. Version 2 adds MIDI support. Version 3 adds VST support.

 

If you don't own a device that can accept MIDI as input and output, you can't use that feature of the software, but you can still use VST plugins, even if they were introduced later. As you can with pretty much all features added later, if they don't require dedicated hardware to work. The tech is always there, you just can't use it because you lack the hardware to do so. Same with DLSS.

[rcmaehl]

Only a couple years until DLSS 3.5E, and then an unpopular DLSS4 followed by a reasonably good DLSS 5 

 

I roll for initiative 

[Eigenvektor]

11 minutes ago, StDragon said:

I should have clarified: Allow me to re-word that.

"So if you have a 20 series, why not call it DLSS 2.5 since the 20 series can't do FG."

Because that would be silly and confusing. It's still an improved version of DLSS version 3 and still adds a new feature on top of DLSS version 3. It can do everything DLSS3 can do + it adds a new feature.

 

I think people here are confusing themselves by thinking of DLSS2 and Super Resolution as being the same thing. They are not. Super Resolution is a feature supported by DLSS. The same way Frame Generation is a feature of DLSS and Ray Reconstruction is a feature of DLSS. They are features introduces in different versions of DLSS. And some of them require additional hardware to be enabled.

 

You're still running DLSS version 3.5 on your 20-series GPU, you simply can't use all of its available features, because you don't have the hardware support for them.

 

Let's say Firefox 3 adds a new feature. But you can only use that feature when your PC has, for example, a TPM chip. Now they introduce Firefox 3.5 which adds yet another new feature that is available to everyone. You don't then label Firefox 3.5 as "Firefox 2.5" on machines that don't have a TPM chip. It's still Firefox 3.5, it still supports all the features of Firefox 3 (+more), you simply can't use all of them due to a lack of hardware.

[StDragon]

28 minutes ago, Eigenvektor said:

Let's say Firefox 3 adds a new feature. But you can only use that feature when your PC has, for example, a TPM chip. Now they introduce Firefox 3.5 which adds yet another new feature that is available to everyone. You don't then label Firefox 3.5 as "Firefox 2.5" on machines that don't have a TPM chip. It's still Firefox 3.5, it still supports all the features of Firefox 3 (+more), you simply can't use all of them due to a lack of hardware.

That is why it's confusing. the concept you can enable 3.5 of whatever, but sub-features may or may not be available depending on the HW.
Technically, I get it. From a marketing perspective, it's stupid. It should be binary yes/no. Yes or no that it can fully support ALL features of 3.5; not some half-assed measure. 

[Eigenvektor]

1 minute ago, StDragon said:

That is why it's confusing. the concept you can enable 3.5 of whatever, but sub-features may or may not be available depending on the HW.
Technically, I get it. From a marketing perspective, it's stupid. It should be binary yes/no. Yes or no that it can fully support ALL features of 3.5; not some half-assed measure. 

Except you don't enable DLSS 3 or 3.5. You do enable its individual features.

 

Look at the screenshots that @tikker attached above. Games don't make you choose between different DLSS versions, they allow you to enable/disable features of DLSS, offered by the particular version they implemented.

 

Without a 40-series GPU some of the toggles are simply off and disabled.

 

When a game implements DLSS version 3.5, you'll simply get an additional toggle for RR. And you'll still have the toggle for FG, that is enabled or disabled based on hardware support. Plus the toggle/slider for SR quality.

[SnugglyGirl]

Does anyone else think that Nvidia is going to start increasing the cycle of GPU's and then focus on DLSS to plug in the 2+ year gap between with great improvements to try and entice people to upgrade to the latest rtx cards?

[starsmine]

3 minutes ago, SnugglyGirl said:

Does anyone else think that Nvidia is going to start increasing the cycle of GPU's and then focus on DLSS to plug in the 2+ year gap between with great improvements to try and entice people to upgrade to the latest rtx cards?

what do you mean increase the cycle of GPUs?

We already know when the next architecture comes out and its "late" in comparison to the old cadence. Lovelace-Next (be that Blackwell or something else) will not be out earlier than spring 2025.

Nvidia can refresh Lovelace next year sure, but I have not run into credible rumors about that.

[SnugglyGirl]

Just now, starsmine said:

what do you mean increase the cycle of GPUs?

We already know when the next architecture comes out and its "late" in comparison to the old cadence. Lovelace-Next (be that Blackwell or something else) will not be out earlier than spring 2025.

Nvidia can refresh Lovelace next year sure, but I have not run into credible rumors about that.

The cycle for GPU's has been slowly increasing. I remember when it was every 6 months -1 year a new GPU was released. Now this generation coming will have been well over 2 years as it is coming out in 2025

[porina]

7 hours ago, SnugglyGirl said:

The cycle for GPU's has been slowly increasing. I remember when it was every 6 months -1 year a new GPU was released. Now this generation coming will have been well over 2 years as it is coming out in 2025

Was it ever as short as 6 months for a full generation? I wouldn't count mid-cycle refreshes in that.

 

Last I saw the rumour was the next gen NV might be early 2025, so perhaps a quarter over 2 years. If that is the case, I don't see it as a major deal. If it were extended to something like 3 years, that would be more of a concern.

[tikker]

8 hours ago, SnugglyGirl said:

The cycle for GPU's has been slowly increasing. I remember when it was every 6 months -1 year a new GPU was released. Now this generation coming will have been well over 2 years as it is coming out in 2025

If they need more time then I think that's fine. It is not just getting longer for the sake of stretching it out. Getting the next big improvement is also just harder compared to 10 years ago.

[GOTSpectrum]

2 hours ago, porina said:

Was it ever as short as 6 months for a full generation? I wouldn't count mid-cycle refreshes in that.

 

Last I saw the rumour was the next gen NV might be early 2025, so perhaps a quarter over 2 years. If that is the case, I don't see it as a major deal. If it were extended to something like 3 years, that would be more of a concern.

Didn't Pascal last like 2-3 years?

[Eigenvektor]

2 hours ago, porina said:

Was it ever as short as 6 months for a full generation? I wouldn't count mid-cycle refreshes in that.

Yes, two decades ago

 

GeForce 2 came out in 2000-2001 and GeForce 3 launched early 2001 with GeForce 4 already launching the following year.

 

The performance gain between generations was also quite a lot back then.

 

But that's simply a sign of a new technology with easy to make gains. Now that GPUs have been around a long time, it takes a lot more work to improve.

[Mihle]

14 hours ago, Eigenvektor said:

Except you don't enable DLSS 3 or 3.5. You do enable its individual features.

 

Look at the screenshots that @tikker attached above. Games don't make you choose between different DLSS versions, they allow you to enable/disable features of DLSS, offered by the particular version they implemented.

 

Without a 40-series GPU some of the toggles are simply off and disabled.

 

When a game implements DLSS version 3.5, you'll simply get an additional toggle for RR. And you'll still have the toggle for FG, that is enabled or disabled based on hardware support. Plus the toggle/slider for SR quality.

I think part of his point is that it should not be marketed as DLSS 3 or 3,5 to consumers but some previous features not supported. It should instead be marketed as something easier to understand for consumers. For example:

DLSS FG/Frame Generation

DLSS SR/Super Resolution

DLSS RR/Ray Reconstruction

 

They can then just say X GPU supports DLSS SR, RR (but not FG) for example.

And that the DLSS 3 or 3.5 is more behind the scenes naming.

[porina]

2 hours ago, GOTSpectrum said:

Didn't Pascal last like 2-3 years?

1080 May 2016

2080 Sep 2018

2 years 4 month gap in that case.

 

1 hour ago, Eigenvektor said:

Yes, two decades ago

 

GeForce 2 came out in 2000-2001 and GeForce 3 launched early 2001 with GeForce 4 already launching the following year.

Found it hard to nail down an exact release date. Wikipedia says mid-May 2000 on the page for GF2, although earliest model on the list page is late April. Could be difference between announce and release? GF3 following up February 2001, which would be about 9 months.

 

I did find this comment from some guy called Anand:

Quote

By this time, the market was just getting used to NVIDIA’s 6-month product cycles, which, when executed properly, would result in a new technology being launched every fall followed by a “spring refresh,” which would boast the move to higher clock speeds and an upgraded feature set to keep the current product generation alive until the fall where NVIDIA would again introduce a new technology.  

https://www.anandtech.com/show/537

 

[Eigenvektor]

6 hours ago, Mihle said:

I think part of his point is that it should not be marketed as DLSS 3 or 3,5 to consumers but some previous features not supported.

Except DLSS in version 3.5 does not support less features than DLSS in version 3. DLSS 3.5 adds one additional feature: Ray Reconstruction. Otherwise, it has the exact same feature set as DLSS 3.

 

I think the issue is that when DLSS was first introduced everybody, myself included, thought of DLSS and Super Resolution as synonyms. Maybe Nvidia should've made the distinction more clear. But they probably didn't even plan on adding more features to DLSS at that time.

 

By now DLSS is a technology stack; it's a collection of features based around machine learning. The first two versions simply only supported Super Resolution.

 

When they introduced DLSS 3, that earlier misconception seems to have lead to the impression that DLSS 3 and Frame Generation are also synonyms. They are not. Frame generation is simply a feature that was added to their existing technology stack under the "DLSS" branding. Who knows, maybe they should've renamed it to Nvidia Machine Learning Technologies at this point and that would've been less confusing.

 

6 hours ago, Mihle said:

as something easier to understand for consumers. For example:

DLSS FG/Frame Generation

DLSS SR/Super Resolution

DLSS RR/Ray Reconstruction

 

They can then just say X GPU supports DLSS SR, RR (but not FG) for example.

And that the DLSS 3 or 3.5 is more behind the scenes naming.

Which is how it is generally presented to the casual end user (e.g. in games). For example here are the settings from Cyberpunk 2077:

 

 

There is zero mention of DLSS 2 or 3. It simply lists the features that are part of Nvidia DLSS and you can enable or disable them. The user doesn't need to know anything about the DLSS version supported by the game. And if one of these features isn't supported due to a lack of hardware support, it's simply grayed out.

 

Once the game is updated to DLSS 3.5, I expect there will simply be another checkbox added for "DLSS Ray Reconstruction".

[tikker]

6 hours ago, Mihle said:

I think part of his point is that it should not be marketed as DLSS 3 or 3,5 to consumers but some previous features not supported. It should instead be marketed as something easier to understand for consumers. For example:

DLSS FG/Frame Generation

DLSS SR/Super Resolution

DLSS RR/Ray Reconstruction

 

They can then just say X GPU supports DLSS SR, RR (but not FG) for example.

And that the DLSS 3 or 3.5 is more behind the scenes naming.

That is what games tend do as you can see in the screenshots on the previous page. General consumers will just see the option appear once games update. Nvidia's news page shows versions 2.3 and 3.1 existed, but if I were to guess those flew under the radar simply because they were relatively uninteresting normal updates that didn't introduce an big new feature like 3.5 does, alongside any other normal updates.

 

I'm not sure if supporting a certain DLSS version forces you to implement all technologies available in that version. The wording surrounding DLSS does imply that to me. If you want a game that supports ray reconstruction you can at least say it needs to at minimum support DLSS 3.5. If Nvidia does indeed enforce that you implement all technologies of a certain version then you can safely say that games supporting DLSS 3.5 will offer ray reconstruction if hardware permits. That should just not be confused with the different statement of saying that DLSS 3.5 is ray reconstruction.

[Mihle]

4 hours ago, Eigenvektor said:

Except DLSS in version 3.5 does not support less features than DLSS in version 3. DLSS 3.5 adds one additional feature: Ray Reconstruction. Otherwise, it has the exact same feature set as DLSS 3.

 

I think the issue is that when DLSS was first introduced everybody, myself included, thought of DLSS and Super Resolution as synonyms. Maybe Nvidia should've made the distinction more clear. But they probably didn't even plan on adding more features to DLSS at that time.

 

By now DLSS is a technology stack; it's a collection of features based around machine learning. The first two versions simply only supported Super Resolution.

 

When they introduced DLSS 3, that earlier misconception seems to have lead to the impression that DLSS 3 and Frame Generation are also synonyms. They are not. Frame generation is simply a feature that was added to their existing technology stack under the "DLSS" branding. Who knows, maybe they should've renamed it to Nvidia Machine Learning Technologies at this point and that would've been less confusing.

 

Which is how it is generally presented to the casual end user (e.g. in games). For example here are the settings from Cyberpunk 2077:

 

 

There is zero mention of DLSS 2 or 3. It simply lists the features that are part of Nvidia DLSS and you can enable or disable them. The user doesn't need to know anything about the DLSS version supported by the game. And if one of these features isn't supported due to a lack of hardware support, it's simply grayed out.

 

Once the game is updated to DLSS 3.5, I expect there will simply be another checkbox added for "DLSS Ray Reconstruction".

 

2 hours ago, tikker said:

*snip*

I know they do that in games settings, but that's not what I talked about. I said the marketing. If you look up lot of the marketing for DLSS, a lot of it just day "DLSS 3 On" and off, not "DLSS Frame Generation on" and off.

[Eigenvektor]

9 hours ago, Mihle said:

I know they do that in games settings, but that's not what I talked about. I said the marketing. If you look up lot of the marketing for DLSS, a lot of it just day "DLSS 3 On" and off, not "DLSS Frame Generation on" and off.

I agree that this type of wording is bad and misleading

 

~edit: Now that I've thought about it some more, you're only partially correct. Because DLSS 3: On and DLSS FG: On are not equivalent.

 

I agree it would be better if they listed individual features, especially with DLSS 3.5, where new features are not limited to new GPUs.

 

But right now DLSS 3: On is essentially shorthand for DLAA: On, DLSS SR: On & DLSS FG: On

[tikker]

20 hours ago, Mihle said:

I know they do that in games settings, but that's not what I talked about. I said the marketing. If you look up lot of the marketing for DLSS, a lot of it just day "DLSS 3 On" and off, not "DLSS Frame Generation on" and off.

I agree that using the version in marketing isn't great. That relates to my other point of whether supporting a certain version means you must implement all the features associated with it, and from what I can tell (partly because of the marketing) that seems to be the case. In that scenario I guess they are technically correct, since SR also gets updates, so it is not just FG on vs FG off. As an entire package it's not the worst in my eyes to then say "here's what DLSS 2 could do, now here's what DLSS 3.5 in full gear can do". For purely showing off e.g. the frame generation feature I agree it would be better if they marketed the comparison of DLSS 3 SR vs DLSS 3 SR+FG.