Discussion Nvidia Blackwell in Q1-2025

[Hulk]

My biggest issue with interpolation as opposed to prediction (although currently I'm kind of "opposed" to both) is that with interpolation you buffer frames before they are displayed. So while you are SEEING smoother frame rates, you are FEELING the exact same lag that you would feel at native frame rates. So honestly, what's the use of interpolation, especially at low native frame rates where you can feel input lag.

I need to know more about this but currently I'm thinking that this might not be the best road to go down. Perhaps it would be better to just focus on improving good old fashioned, honest rendering of frame. Work on improving the hardware and the software instead of strange workarounds.

 

[Bryo4321]

IDK, I would think that such an educated and experienced industry veteran might be less biased but IG not.

Software engineers (especially the talented ones) tend to be strongly opinionated about the right way to do things. I doubt it’s related to allegiance to one company or the other and more which aligns with what he thinks is the right way to do things.

Having worked with many that is my experience, so that’s my guess.

 

[ajsdkflsdjfio]

You don't combine fallacies like that and no, they are separate things. If both exist in an argument, you expose them individually. The rest of it is just saying a lot without saying much (and yes he does say having motion artifacts in movie theaters is unacceptable, it's a rhetorical question).

I'll make it simple.

He doesn't like the artifacts from frame generation and doesn't think the industry should be going down this path. He doesn't think we should accept the visual artifacts and issues with frame generation in high fidelity visual media. That's it, simple. I happen to agree with him. He actually gave some props to the improvements in DLSS4 in terms of upscaling (though maybe a little back handed, hard to tell tone) so he seems to be in favor of upscaling in general but against frame generation.

Who is combining fallacies here. Straw man arguments often utilizes some for of false equivalency, if you can't understand that then there is no need for us to talk anymore. I wasn't trying to combine fallacies, I was merely breaking down the multiple faults in his argument, one part of which was drawing a false equivalency between the role of visual fidelity in a Movie Threatre vs PC gaming.

Definition of strawman:

Exaggerating an opponent's argument:
Opponents argument: certain amounts of upscaling/MFG artifacts are okay in PC graphics
Exaggeration: certain amounts of upscaling/MFG artifacts are okay in Movie Theatre screenings
Attacking the exaggerated version: certain amounts of upscaling/MFG are NOT okay in Movie threatre screenings.

You obviously disagree, so rather than try to spend a thousand words dismissing someone's opinion, why not simply disagree and say why you think the artifacts are worth what you consider to be the advantages of the technology? You'd save everyone a ton of time and be arguing in good faith versus trying to appeal to censure and coming across as overly emotional about a rendering technology.

My original post literally did exactly what you said, I simply disagreed and stated that in my personal opinion upscaling technology is useful and the artifacts aren't distracting enough to warrant a complete writing off of it's usefulness. I wrote a singular message pointing out that I felt that he was a bit too critical of DLSS not just of it's flaws, but of it's actual benefits for consumers. All this other discussion is merely me responding to people like you who didn't like me calling out the flaws in his argument.

Don't be mistaken here, much of his discussion is about upscaling too, in the first place around half of his screenshots are regarding upscaling artifacts, and his discussion about AI never being able to reach the correct answer is in direct response to Nvidia's upscaling advances with their shift from a CNN to Transformer model.

 

[Heartbreaker]

My biggest issue with interpolation as opposed to prediction (although currently I'm kind of "opposed" to both) is that with interpolation you buffer frames before they are displayed. So while you are SEEING smoother frame rates, you are FEELING the exact same lag that you would feel at native frame rates. So honestly, what's the use of interpolation, especially at low native frame rates where you can feel input lag.

I need to know more about this but currently I'm thinking that this might not be the best road to go down. Perhaps it would be better to just focus on improving good old fashioned, honest rendering of frame. Work on improving the hardware and the software instead of strange workarounds.

You will feel even worse lag.

But it has uses smooth out the video on screen. If you have a very high refresh monitor, it can significantly improve clarity, by reducing the Sample and Hold effect. So it's not totally useless.

But I strenuously object to how it's marketed as real FPS and comparing to other cards running without FG or less MFG, to make new cards appear faster than they really are.

Market it's video smoothing, or clarity enhancing and I wouldn't have an issue.

Using it for deception is the problem, not the technology itself.

 

[Hitman928]

Who is combining fallacies here.

You are.

Straw man arguments often utilizes some for of false equivalency, if you can't understand that then there is no need for us to talk anymore.

If you say so.

I wasn't trying to combine fallacies, I was merely breaking down the multiple faults in his argument, one part of which was drawing a false equivalency between the role of visual fidelity in a Movie Threatre vs PC gaming.

Definition of strawman:
View attachment 114624
Exaggerating an opponent's argument:
Opponents argument: certain amounts of upscaling/MFG artifacts are okay in PC graphics

Exaggerating, oversimplifying, creating arguments that are tangents off of theirs, all straw men.

Exaggeration: certain amounts of upscaling/MFG artifacts are okay in Movie Theatre screenings

No one did this.

Attacking the exaggerated version: certain amounts of upscaling/MFG are NOT okay in Movie threatre screenings.

You seem to not follow the argument.

My original post literally did exactly what you said, I simply disagreed and stated that in my personal opinion upscaling technology is useful and the artifacts aren't distracting enough to warrant a complete writing off of it's usefulness. I wrote a singular message pointing out that I felt that he was a bit too critical of DLSS not just of it's flaws, but of it's actual benefits for consumers. All this other discussion is merely me responding to people like you who didn't like me calling out the flaws in his argument.

Your first one wasn't too bad but then you seemed to go off the rails when people agreed with his assessment and comparison. I don't even care if you like it or not, I just hate people appealing to censure, especially in an open forum; unchecked it can quickly devolve into a WCCFtech experience. I won't reply on this anymore since it's just cluttering the forum, especially when each of your posts are like 5 from other people. I'd much rather just discuss the technology itself.

 

[Hitman928]

You will feel even worse lag.

But it has uses smooth out the video on screen. If you have a very high refresh monitor, it can significantly improve clarity, by reducing the Sample and Hold effect. So it's not totally useless.

But I strenuously object to how it's marketed as real FPS and comparing to other cards running without FG or less MFG, to make new cards appear faster than they really are.

Market it's video smoothing, or clarity enhancing and I wouldn't have an issue.

Using it for deception is the problem, not the technology itself.

I largely agree with this, especially the deceptive marketing bit, but for me, I don't really like the technology itself even. I've tried it on a few games and always end up turning it off. It was a horrible experience playing with it when the actual frame rate is low and only OK when it is high. It feels like playing on a TV with a bunch of post processing where there is a disconnect between inputs and what happens on the screen. Also, when the actual frame rate is already high, the improved motion fluidity is less impactful and you still get some weird artifacts that are disruptive to the immersive feeling of the game. Other people may not mind the added latency as much or not be able to see the artifacts I guess, so I'm sure there are some that like it. If they get a few more versions out that improve the experience, I'll try it again.

 

[poke01]

There show problems with MFG in DF video. Ofc DF said it's nice etc

Look I get that you dislike Nvidia but I won’t watch his videos. One can do same for FSR4 and pixel peep issues.

BTW contrast his thread bashing DLSS4 versus his reaction to FSR4. Ofc there was no FG involved in the FSR4 demo but much of his DLSS4 thread was in fact criticizing upscaling, not MFG. He has worked at both AMD and Nvidia, dk why he seems to favor one of the other so much.

https://twitter.com/x/status/1877351010437378401

that’s why I don’t like those people and don’t post them here. But Pro AMD people post randos here and it’s clear those twitter AMD folks have an anti-Nvidia agenda

 

[Heartbreaker]

I largely agree with this, especially the deceptive marketing bit, but for me, I don't really like the technology itself even. I've tried it on a few games and always end up turning it off. It was a horrible experience playing with it when the actual frame rate is low and only OK when it is high. It feels like playing on a TV with a bunch of post processing where there is a disconnect between inputs and what happens on the screen. Also, when the actual frame rate is already high, the improved motion fluidity is less impactful and you still get some weird artifacts that are disruptive to the immersive feeling of the game. Other people may not mind the added latency as much or not be able to see the artifacts I guess, so I'm sure there are some that like it. If they get a few more versions out that improve the experience, I'll try it again.

It's benefits are mostly theoretical for me. I'm a bit skeptical of even real, super high frame rate stuff, outside of competitive esports gamers.

I can barely tell the difference between good 60 FPS (I notice frame drops right away if Vsycned) and 120 FPS so maybe the benefits are just lost on me.

I'd have to see more examples and maybe get a higher refresh monitor (165Hz max right now).

I only tried FG for a minute or two in Witcher 3 and didn't notice any benefit except a higher frame counter. But as I said before, I barely notice a benefit from real higher FPS, so how could fake higher FPS be better? But again, that could just be me.

I'll test more whenever I get more into the game and am more attuned to its feel. The controls still feel awkward to me, so it's hard to notice extra latency.

It could have a benefit for some people in some situations, so I don't want to dump on the tech itself, just the vile marketing.

 

[poke01]

I don’t watch YouTubers with double standards and I HATE MFG technologies so I’m not just on Vex but any YouTuber that does this for clicks

 

[beginner99]

how they could use prediction and avoid the jarring situations when objects become visible just after the last fully rendered frame.

This means the FPS is massively increasing but the latency is still going up.

They are not extrapolating but interpolating. frame 1 gets rendered, they hold it till frame 2 comes, then in between these 2 frame 3 fake frames are made. Frame 1 only gets shown when frame 2 is available, this means latency always goes up and the lower the frame rate is, the higher the latency increase. So when you actually need frame gen say with 20 fps, it doesn't work very well in any non-slow paces game very latency matters.

This means objects popping in aren't an issue due to interpolation.

it's possible but unlikely dlss 4 will do extrapolation, eg not waiting for frame 2. and then yes, things can get very, very messy as the future can't be predicted like sudden player movements. for some games it could work somewhat with a trained model for that game but then it's game specific again and for shooters with lot's of jerky and unpredictable moevement, I don't see it ever working so yes, 99% it will be interpolation again, that is "easy", just fill the gaps. And >90% of gamers don't notice the latency anyway, eg don't play competitive shooters or RTS.

 

[coercitiv]

They are not extrapolating but interpolating.

In the post you replied to I was exploring a claim in the thread that they might actually use extrapolation. I was expressing doubt.

 

[Gideon]

If it all works well is it fair to say 5070 has a 4090 experience? I don't think 4090 performance wording can be justified. But I wonder if perception is all that matters.

In some cases .... maybe? But definitely not on average.

• If the raw FPS is at or below 30FPS on 5070 and it's 60 on the 4090, then the image might seem equally smooth, but the latency hit would be very noticable
• Also due-to only having half the memory of the 4090 there will definitely be scenarios where 5070 will just run out of memory and fall flat on its face having singile digit FPS and huge uneven spikes on the frame graph. You can't framegen yourself out of that situation.

The RTX 4070 == RTX 3090 Ti comparison was more understandable cause at 1440p, particularily wth DLSS on, RTX 4070 really was quite close, even in rasterization (4070 SUPER is pretty much 3090 or rather 3080 Ti 12GB, once you factor in memory). But 3090 was hardly a step up from the 3080 series. 4090 is massively better. In raw performance. In 4K RT titles it doubles up 4070 Ti (which in raw performance should be within 10% of the 5070) and it can also do framegen, though yeah not 3 fake frames.

At 280 FPS (with framegen) you might get awa with the claim that they offer similar experience. You sure as hell can't do that when the FPS (with framegen) is around 60 FPS as it is in AW2 with path tracing on a 4090.

In fact I played through that entire game with 4090 (loaned that for the very reason) using path tracing at 1440p with DLSS Quality + framegen (max playable settings really for the card). It was very playable (very similar to this experience), but there were definitely performance drops in demanding areas that you most certainly noticed in much worse input lag even when the output was "smoothed over" by framegen. And that was with the AVG FPS being over 100 (with framegen).

TL;DR: the game was only decently playable at settings where the RAW prerformance before framegen was consistently 60-70 FPS. It sure as hell won't be playable when the raw performance (with upscaling) is in the 30ies.

 

[basix]

Regarding Frame Generation it is clear that we have only scratched the surface.

The holy grail of FG would be following: Frame Extrapolation + MFG + Frame Warp

• Frame Extrapolation
  • Not latency increase due to FG
• Multi Frame Generation (MFG): Maybe up to 4x (technology / quality restrictions, especially with extrapolation because that is harder than interpolation)
  • Usability depends on frame rate as there are pronably limits to the size of the extrapolation window, where you can achieve decent quality
    • 30fps -> 60fps -> extrapolation window = 1*33ms = 33ms
    • 40fps -> 120fps -> extrapolation window = 2*25ms = 50ms -> More temporal info allows for a wider extrapolation window -> PS6 and co. with 120fps target?
    • 60fps -> 240fps -> extrapolation window = 3* 16.7ms = 50ms
• Frame Warp Technology (FWT, see Reflex 2)
  • Adjust image to the user input at latest possible time
  • Apply FWT to rendered as well as FG / MFG frames
    • In combination with extrapolation FG / MFG frames will feel like rendered frames regarding latency and gameplay responsivness -> a clear winner
    • If using Reflex / AntiLag with Frame Warp the feeled latency will be drastically reduced already, so if you play at 240fps (60fps rendered) it would feel even better than todays 240fps with Reflex 1 and no FG!

The list above solves all the responsivnes issues of FG. I assume we reach that in a few years of HW, SW and ML/AI development.
The critical point here will be quality and artifacts. When I see how DLSS4 improves Upsampling and Ray Reconstruction quality I have good hopes, that this will also happen to FG and Frame Warp.

 

[Win2012R2]

Looks like in the future we'll get very high refresh OLED displayed that support 640x480 native res as it would be the only way for people who like native rendering to place AAAAA titles...

 

[Gideon]

Looks like in the future we'll get very high refresh OLED displayed that support 640x480 native res as it would be the only way for people who like native rendering to place AAAAA titles...

There already are monitors that can either do FHD @ 480 Hz or UHD @ 240 Hz:

LG Display to Mass Produce World's First Gaming OLED Panel With Switchable Refresh Rate and Resolution

LG Display, the world's leading innovator of display technologies, announced today that it has developed the first-ever Gaming OLED panel with a switchable refresh rate and resolution. The company has begun mass production of the 31.5-inch panel this month, accelerating its push into the...

www.techpowerup.com

4K monitors can do lossless scaling to 1080p and 1440p monitors to 720p so this might actually become a thing

 

[CakeMonster]

What would actually be useful right now is two things:

1) Software that shows the base frame rate when FG is in use, for reviews, and for consumers to get an idea what their baseline is, and then be able to turn off FG if its not really needed.

2) An adaptive FG mode that only kicks in when the frame rate goes low enough to ameliorate those CPU limited scenarios that we are only going to see more of. This is obviously mostly useful for those who can manage to run 100+ fps without FG, and then have it kick in only when it drops below that.

My idea for 2) is based on the adaptive DLSS mode that exists in HFW. It can run DLSS and drops the input resolution to try to reach a frame rate target. So alternating between Quality and Performance for a more stable frame rate. Hopefully they add DLAA to this in the future. This is much preferable to FG IMO, but when CPU limited it simply can't keep up.

 

[SiliconFly]

Looks like in the future we'll get very high refresh OLED displayed that support 640x480 native res as it would be the only way for people who like native rendering to place AAAAA titles...

Never underestimate 640x480!

 

[coercitiv]

2) An adaptive FG mode that only kicks in when the frame rate goes low enough to ameliorate those CPU limited scenarios that we are only going to see more of. This is obviously mostly useful for those who can manage to run 100+ fps without FG, and then have it kick in only when it drops below that.

My idea for 2) is based on the adaptive DLSS mode that exists in HFW. It can run DLSS and drops the input resolution to try to reach a frame rate target. So alternating between Quality and Performance for a more stable frame rate. Hopefully they add DLAA to this in the future. This is much preferable to FG IMO, but when CPU limited it simply can't keep up.

SCC: The Shader Compilation Compensator

Are you tired of frame drops interrupting your gaming experience when shaders are being compiled? Experience smooth gameplay like never before with SCC - Shader Compilation Compensator, the innovative solution designed to keep your frames flowing even under heavy CPU loads. This ensures uninterrupted fluidity, so you can dive deeper into your gaming adventures without missing a beat.

Disclaimer:

The Shader Compilation Compensator (SCC) is designed as an innovative solution to enhance frame generation during periods when your system CPU is taxed by shader compilation. While SCC offers significant improvements in maintaining frame rate continuity under such conditions, it is important to understand its limitations:

• Latency Spikes: SCC is not equipped to address or compensate for latency spikes that may occur due to factors beyond shader compilation, such as network delays, system resource constraints unrelated to graphics processing, or hardware performance issues.
• Temporary Loss of Character Control: During moments when your system experiences temporary control loss—often caused by severe CPU/GPU lag, insufficient system resources, or external interruptions—SCC cannot guarantee the prevention or compensation for these disruptions in gameplay.
• External Factors: The effectiveness of SCC can be influenced by various elements outside its scope, including but not limited to system hardware limitations, driver compatibility issues, and other software running concurrently that may impact overall performance.
• Game-Specific Variability: Different games have unique demands on system resources. As such, the benefits provided by SCC may vary depending on specific game requirements and settings.

 

[SolidQ]

Wukong Path Tracing

 

[DAPUNISHER]

Wukong Path Tracing

Explain it to me like I'm 5

 

[Hitman928]

Explain it to me like I'm 5

At native resolution, Wukong with path tracing is getting 29 fps (unknown native res and card being tested but probably 4K resolution). With DLSS4 upscaling and multi-frame generation enabled, it is getting 244 fps (unknown render resolution but probably 1080p render resolution to 4K display). Because of the low render res and frame gen, there are probably significant visual artifacts and who knows how it actually feels to control, but who cares about that when the fps number displayed goes to the moon!

 

[DAPUNISHER]

At native resolution, Wukong with path tracing is getting 29 fps (unknown native res and card being tested but probably 4K resolution). With DLSS4 upscaling and multi-frame generation enabled, it is getting 244 fps (unknown render resolution but probably 1080p render resolution to 4K display). Because of the low render res and frame gen, there are probably significant visual artifacts and who knows how it actually feels to control, but who cares about that when the fps number displayed goes to the moon!

Thanks man.

I presume the reason the images are different is due to the 29fps being behind on rendering those fake frames?

 

[Hitman928]

Thanks man.

I presume the reason the images are different is due to the 29fps being behind on rendering those fake frames?

I think you are just seeing the top right corner of one and the top left corner of the other. They used different display corners of the FPS counter so that people could compare the fps easily (stare at the fps numbers, pay no attention to the actual quality of the images ).

 

[DAPUNISHER]

I think you are just seeing the top right corner of one and the top left corner of the other. They changed the display corner of the FPS counter so that people could compare the fps easily (stare at the fps numbers, pay no attention to the actual quality of the images ).

Thanks, I hate it. Had me staring for a good minute doing a compare and contrast. I am sitting thinking, is homie pointing out issues with the lighting? The last SS was of a cup turning dark as the fake frames increased. I thought it was another one of those. Turns out it was just an inane SS showing the fake frame rate that has been plastered everywhere already. SMDH.

 

[SiliconFly]

My biggest issue with interpolation as opposed to prediction (although currently I'm kind of "opposed" to both) is that with interpolation you buffer frames before they are displayed. So while you are SEEING smoother frame rates, you are FEELING the exact same lag that you would feel at native frame rates. So honestly, what's the use of interpolation, especially at low native frame rates where you can feel input lag.

With frame interpolation, the latency actually goes up compared to native rendering. For example, in pure native, if a game is running at 30fps, a real frame is displayed every 33ms immediately after rendering without any delay. But once we enable frame interpolation, the real frame gets buffered and is displayed "only after" the fake frame is displayed. It does improve visuals but adds lot of latency.

On the contrary, frame prediction doesn't need to buffer a real frame, and hence can display the real frame immediately after rendering without any delay. The fake frames are produced later, properly paced and displayed. Reduces latency compared to frame interpolation.

Edit: (Note: I maybe wrong about frame prediction cos' I'm assuming here Jensen was being honest and not deceptive when he said 50 series is using AI models to predict forward instead of "buffering" and interpolating like CNNs.)

Even if the generated frames are instantly produced, they still need to be paced. If the system generates one fully rendered frame every 20ms, then 3 extra frames would have to be paced every 5ms to have proper cadence. However, if the cost of generated frames is non-zero, let's say 3ms per frame, then that cost adds to the latency of the system, as fully rendered frames are now generated every ~30ms. (20+3+3+3)

Yes it does add latency when we enable 3X FG. The real frames, instead of being presented every 20ms, are now presented every 30ms, but are done so immediately after rendering is complete without any further delay (that is usually associated with buffering the real frame for interpolation). This should give frame prediction a good advantage in latency over frame interpolation.

Assuming all things they said during the presentation were true, 5090 (with 3X FG) should have lesser latency than 4090 (with 1X FG).

The following frame pacing is far from reality. But in an ideal world, this is how I expect it to work with frame prediction. Just the idea: