DLSS 5 preview

[bingham67]

 

[CAT-THE-FIFTH]

10ms is the minimum, that's with a 5090, the lower down the GPU tier you go the worse it gets and when you add frame gen on top of that try and get some FPS back the latency starts to add up to something that feels really laggy.

All of these technologies have their own render time that gets added and stacked on top of your frame render time and eachother.

Well unlike the others here I have a mainstream Nvidia card in my other system,so if people want me to test some games on pleb hardware I can do so.

 

[SkeeterUK]

Wonder if it works for tdu 2.

 

[Neil79]

 

[Calin Banc]

Plenty of artifacts there...

 

[Rroff]

Wonder if it works for tdu 2.

TDU2 is a bit of a mess to mod with the DRM and they've done some tricks with the renderer which don't play nice with indiscriminately chucking the game at a different rendering pipeline I had it loaded up in Remix to mess about with changing the lighting and textures and it ran at like 5 FPS

Though it probably works as well as anything else is so far if you use the desktop neural renderer.

EDIT: There are videos of TDU 2 with DLSS 5 - though results are not amazing.

 

[Dicehunter]

Quite impressed with Assassins Creed Shadows, Using the Neural Upstream method with the highest quality preset, Image quality stays the same as the normal DLSS 5 method but performance is massively improved, With DLSS at performance and basic 2x FG I'm seeing around 120FPS at 5120x2160, Everything maxed out and power use hovers around 350w, I imagine this is likely what most newer games will run like where DLSS 5 can more easily do its thing where as games where we see 600w and bad performance it is more along the lines of brute force.

Really nicely enhances the environment too.

 

[KompuKare]

Well unlike the others here I have a mainstream Nvidia card in my other system,so if people want me to test some games on pleb hardware I can do so.

Blasphemy!
Only 5090 owners are allowed to try this!

 

[TimTamMan]

Hah, as I commented earlier, modders move way faster and with mind better result in this case than the multi billion corporation. Not very surprising these days - Nvidia CEO himself said many time AI at home will speed up development of many things. Just be didn't realise that would make his own company just look bad

I find it quite funny that the LLMs that are trained on Nvidia GPUs are being used to erode software advantage Nvidia has over other producers

 

[Kirby Wurm]

Is it just me, or do a good percentage of the character faces with DLSS5 enabled look like they're under the same lighting conditions, regardless of the game?

What do non-human character faces look like? I've yet to try it as all I've been playing for the last 3 months is Escape From Tarkov PvE.

 

[Calin Banc]

I disagree with this.

The reason why we are going down this path of AI is because moores law is indeed dead. You will not see exponentially more powerful hardware able to run games at this same level of detail using simply compute.

This is why instead of relying on pure compute they are going down the AI route which is sort of like a software emulation level which can accelerate much faster than the lagging hardware advances.

So basically moores law in this instance has moved from hardware to the software side where we will see continued doubling of "emulated compute" every 2 years.

Do you really think that if Devs coded their games to be more detailed to use up this 10ms buffer we would be able to run the games at all with current gen or even near future gen hardware?

The only reason why we are having this conversation is because it's here only by virtue of DLSS5 and AI.

Even now a 5080 would struggle with a lot of games run full RT and PT at 4k with no upscaling.

Why do you think it's a good idea to add an additional 10ms of computation to this equation and think we can still carry on down the hardware route?

You've got to pick your battles...

5080 4k dlss performance with RT Reflections in NBA 2027 is 102 fps vs dlls 5 with 43fps... Basically you're left with 42% of the initial performance.

4080 4k dlss performance with RT Reflections in CB 77 is 79fps vs 52 with path tracing. That's 66% of the initial performance by going to path tracing! Raster is 108fps... so path tracing vs raster (with reflections to high, which offers more performance than psycho) is around 48%. Ergo, going from raster to path tracing is faster than going from whatever to dlss5... And path tracing will always be superior.

The Callisto Protocol has one of the best (natural) looking characters that I've seen in games, both in materials quality and animations PLUS is fast.

Is it just me, or do a good percentage of the character faces with DLSS5 enabled look like they're under the same lightning conditions, regardless of the game?

What do you non-human character faces look like? I've yet to try it as all I've been playing for the last 3 months is Escape From Tarkov PvE.

It appears the same to me (and also a bit of Clarity and / or Dehaze abuse at times... those that use Lightroom will know ) and that's most likely due to its training. In the end... it looks off.

 

[opethdisciple]

You've got to pick your battles...

5080 4k dlss performance with RT Reflections in NBA 2027 is 102 fps vs dlls 5 with 43fps... Basically you're left with 42% of the initial performance.

4080 4k dlss performance with RT Reflections in CB 77 is 79fps vs 52 with path tracing. That's 66% of the initial performance by going to path tracing! Raster is 108fps... so path tracing vs raster (with reflections to high, which offers more performance than psycho) is around 48%. Ergo, going from raster to path tracing is faster than going from whatever to dlss5... And path tracing will always be superior.

The Callisto Protocol has one of the best (natural) looking characters that I've seen in games, both in materials quality and animations PLUS is fast.



It appears the same to me (and also a bit of Clarity and / or Dehaze abuse at times... those that use Lightroom will know ) and that's most likely due to its training. In the end... it looks off.

I'm not sure what point your trying to make. The level of realism offered by DLSS 5 is not currently doable using native GPU compute which is why we are doing it with AI.

Also dont forget that its an option. You can turn it off if you want.

But what I said about Moores law being dead is a fact.

If we weren't going down this road of AI, graphics would be stagnating along side the slow progress that GPU hardware is making.

Even now we have to rely on DLSS for upscaling just to play games at 4k as the hardware is not there to run it natively.

With out DLSS we wouldn't have RT or PT at all as it just wouldn't be feasible to run.

Remember that this technology will advance performance wise very quickly.

NVIDIA already claims to have improved DLSS 5 performance by over five times (5X or 500%) over a six-month period and I imagine the next gen of cards will probably have some sort of hardware level improvements to drive this sort of thing even faster.

So the balance of progress has shifted more in to the realms of software advancements rather than hardware. We are still seeing hardware advancement, but it is just slower and will probably keep getting slower until there is some sort of big break through in hardware technology that changes the game.

Progress is going to be in the software side. I.e AI.

---

Also to take your example

"5080 4k dlss performance with RT Reflections in NBA 2027 is 102 fps vs dlls 5 with 43fps... Basically you're left with 42% of the initial performance"

OK, that may be so but what about if in 6 months time its only 25% reduction in fps?

Its like enabling RT. It kills your fps. Yes it does. But it also enhances the image quality.

I guarantee you DLSS 5.5 will be even faster with better IQ.

---

I also regard DLSS 5 as still preview technology with the NBA game just a tech demo of the technology Nvidia have clearly sponsored to show it off and get some real world performance metrics to measure.

 

[SkeeterUK]

Someone after a bit of fiddling got it working on detroit game which is a pretty good looking game and looks good in that. I wonder if the syberia games could use it as well. The latest ones dont look too bad for the characters.

Someone try it on the last monkey island game with the weird art style for the lolz.

 

[Noxia]

I find it quite funny that the LLMs that are trained on Nvidia GPUs are being used to erode software advantage Nvidia has over other producers

Using one atm to tell me which previously banned member you are.

 

[CAT-THE-FIFTH]

Blasphemy!
Only 5090 owners are allowed to try this!

At least I will get a feel whether this is actually workable for mere mortals!

 

[TimTamMan]

Using one atm to tell me which previously banned member you are.

I hope results do not disappoint

 

[Tinek]

I disagree with this.

The reason why we are going down this path of AI is because moores law is indeed dead.

I'll cut you right here - that would be correct if it improved performance but it doesn't. And with traditional methods you can do just as well (as some games proven) and it would likely cost less performance hit. In addition, Huawei just presented new phone with folded logic design, which they spoken about earlier this year and everyone assumed it's just marketing bs and the only way to increase performance is to go into much more expensive lower nm. It turned out not only not marketing bs but they already have that tech - same old process and yet new chip over 42% more performance with just 7% higher heat generation and yet easier to cool and much higher transistor density. In other words, they just reanimated progress by doing something that the west considered too expensive and too hard to do it seems (but likely it was just about money as always).

Do you really think that if Devs coded their games to be more detailed to use up this 10ms buffer we would be able to run the games at all with current gen or even near future gen hardware?

And yet DLSS5 adds that min.10ms and people seem happy about it? What are you even talking about? You seem to be contradicting yourself.

 

[Kirby Wurm]

It appears the same to me (and also a bit of Clarity and / or Dehaze abuse at times... those that use Lightroom will know ) and that's most likely due to its training. In the end... it looks off.

To be fair that screenshot looks better as the lighting is very different and it does show.

Also, reading back my own nobile-induced typos in your quote of my post was horrendous . Edited it now!

 

[Dicehunter]

What do non-human character faces look like?

Here you go, Star Wars Outlaws

 

[opethdisciple]

I'll cut you right here - that would be correct if it improved performance but it doesn't. And with traditional methods you can do just as well (as some games proven) and it would likely cost less performance hit. In addition, Huawei just presented new phone with folded logic design, which they spoken about earlier this year and everyone assumed it's just marketing bs and the only way to increase performance is to go into much more expensive lower nm. It turned out not only not marketing bs but they already have that tech - same old process and yet new chip over 42% more performance with just 7% higher heat generation and yet easier to cool and much higher transistor density. In other words, they just reanimated progress by doing something that the west considered too expensive and too hard to do it seems (but likely it was just about money as always).

And yet DLSS5 adds that min.10ms and people seem happy about it? What are you even talking about? You seem to be contradicting yourself.

Are you insinuating then that Nvidia are purposely holding back hardware advancement to push this AI narrative?

Because I will remind you that it not just Nvidia that has to use AI upscalers to get acceptable frame rates. AMD, Intel and Sony all have their own AI upscalers so what does that tell you about the state of pc hardware if the hardware is there to do all the graphics processing natively?

I asked AI this question:

Is the use of technology such as DLSS 5 a way to get further enhancement in the GPU space despite moores law being dead?

Answer:

Yes, technologies like DLSS 5 are the exact definition of how the industry is bypassing the death of traditional Moore's Law. However, DLSS 5 shifts the paradigm entirely compared to previous versions. [1]
While older iterations of DLSS were designed to save hardware resources, DLSS 5 is an aggressive pivot into Neural Rendering. It replaces raw brute-force computing with advanced AI algorithms to achieve next-generation visuals. [1, 2, 3]

---

How DLSS 5 Outsmarts Moore's Law
Instead of relying on the physical shrinking of transistors to squeeze out more performance, Nvidia is leaning into software and hardware co-design: [1, 2]

• 3D-Guided Neural Rendering: Traditional rendering requires calculating millions of complex light bounces (ray tracing), which is incredibly heavy on standard silicon. DLSS 5 bypasses this by using an AI diffusion model. It takes a single rendered frame, reads the motion vectors, and dynamically "repaints" photorealistic lighting, subsurface skin scattering, and material textures onto the scene in real time. [1, 2, 3, 4]
• Algorithmic Efficiency over Transistor Count: By letting dedicated Tensor cores handle the heavy lifting of visual reconstruction, the primary graphics pipeline can focus on basic geometry. The AI fills in the cinematic gaps that hardware can no longer cheapen. [1, 2, 3]

The Twist: The "Moore's Law" Tax is Still Lurking
While DLSS 5 successfully unlocks movie-quality graphics that traditional silicon couldn't natively render, its launch has exposed a new bottleneck: it is exceptionally heavy to run. [1, 2, 3, 4]

• A Heavy Performance Penalty: Unlike previous DLSS versions that acted as a performance "cheat code" to boost frame rates, DLSS 5 causes a massive 50% to 70% drop in frame rates when turned on. It requires so much compute power that early tech demos required two flagship graphics cards to run smoothly. [1, 2, 3]
• The Need for Architecture Innovations: Because it is so resource-intensive, Nvidia and modders are already rushing to optimize it. They are attempting to port the model to the RTX 50-series' new 4-bit floating-point format (NVFP4), which reduces the memory and compute footprint on Blackwell architecture. [1]

The Bottom Line
DLSS 5 proves that enhancement in the GPU space is no longer just about adding more transistors to a chip. It is about using generative AI as a tool to bridge the physical limits of hardware. However, because Moore's Law is dead, optimizing these heavy AI models to run efficiently on a single consumer GPU has become the new engineering battleground. [1, 2, 3, 4]


---

Note this part:

While DLSS 5 successfully unlocks movie-quality graphics that traditional silicon couldn't natively render.