NVIDIA's Frame Generation Technology Could Come to GeForce RTX 30 Series

[AusWolf]

RDNA doesnt actually have ai cores though (matrix).

It has something similar, but I'll look into that.

Yes, I think it is.

So you think Nvidia spends the vast majority of its marketing time and resources on a feature that literally nobody cares about? C'mon...

I'm not even entirely sure what you are arguing with here. That if nvidia tried, they could somehow make it work on ampere and Turing? Sure they could, I have no doubt about it. The question is, why should they spend resources doing that for free, and what would the end result be? Would it actually work properly? And if it didn't, we would go back to the same argument about artificial limitations and nvidia being greedy making it work like crap.

1. Caring about existing customers = better company image.
2. Wider game adoption.
3. Even if it doesn't work properly, it's at least a demo for your future purchase, just like RT on a 1660 Ti was.
4. If spending resources on it doesn't make sense, then why did they decide to do it now? Why not just leave it?

 

[JustBenching]

It has something similar, but I'll look into that.


So you think Nvidia spends the vast majority of its marketing time and resources on a feature that literally nobody cares about? C'mon...


1. Caring about existing customers = better company image.
2. Wider game adoption.
3. Even if it doesn't work properly, it's at least a demo for your future purchase, just like RT on a 1660 Ti was.
4. If spending resources on it doesn't make sense, then why did they decide to do it now? Why not just leave it?

They didn't decide to do it now. They are using a different model now. The old one still doesn't work on older cards, and the new one (transformer also includes dlss) might end up painfully slow, but we have to see.

 

[AusWolf]

They didn't decide to do it now. They are using a different model now. The old one still doesn't work on older cards, and the new one (transformer also includes dlss) might end up painfully slow, but we have to see.

A different model. Right. Like they couldn't use, or didn't think of using that different model right from the beginning because of... reasons.

 

[JustBenching]

A different model. Right. Like they couldn't use, or didn't think of using that different model right from the beginning because of... reasons.

Because it's slower. Which is something we can actually test since we will have access to both.

It's not the first time they changed it btw, dlss 1 was completely different to dlss 2 as well.

 

[AusWolf]

Because it's slower. Which is something we can actually test since we will have access to both.

It's not the first time they changed it btw, dlss 1 was completely different to dlss 2 as well.

That shouldn't have prevented them from adding it to RTX 30, just like they added RT for the 1060 6 GB and up just for show.

 

[dyonoctis]

Does Nvidia? As far as I know, there are no technical details available on Tensor cores, DLSS or FG. Nobody except for an Nvidia engineer knows how they work. Therefore, I believe what I want, and I'd rather not put my trust in a company whose goal is making profit, whether it's called Nvidia, Intel or AMD.

And that's where you start to appear to have a bit of bad faith here, since it doesn't look like you actually try to look for that data...
NVIDIA-Turing-Architecture-Whitepaper.pdf
NVIDIA A100 Tensor Core GPU: Performance and Innovation | IEEE Journals & Magazine | IEEE Xplore
nvidia-ampere-architecture-whitepaper.pdf
nvidia-ada-gpu-science.pdf
1803.04014

 

[AusWolf]

And that's where you start to appear to have a bit of bad faith here, since it doesn't look like you actually try to look for that data...
NVIDIA-Turing-Architecture-Whitepaper.pdf
NVIDIA A100 Tensor Core GPU: Performance and Innovation | IEEE Journals & Magazine | IEEE Xplore
nvidia-ampere-architecture-whitepaper.pdf
nvidia-ada-gpu-science.pdf
1803.04014
View attachment 380828

With all due respect, I shouldn't be required to read an 86-page long whitepaper full of programming jargon that, not being a programmer, I'll never understand, or to subscribe to an online journal just to know what a Tensor core is, or how it's connected to DLSS and FG. Reviews always divulge architectural details in layman's terms to hobby builders like myself, but such easily comprehensible info on why the optical flow thingy (is) was required for Nvidia's FG, and why it isn't for AMD's is awkwardly missing.

The TPU review of the 4090 for example said that the optical flow thingy is required for FG, that's why the feature is missing in Ampere and Turing. Then an article came sometime later stating that the OFA has always been part of Ampere as well as Turing, it just... ehm... (*nvidia PR person searching through notes*)... isn't fast enough.

 

[Godrilla]

With all due respect, I shouldn't be required to read an 86-page long whitepaper full of programming jargon that, not being a programmer, I'll never understand, or to subscribe to an online journal just to know what a Tensor core is, or how it's connected to DLSS and FG. Reviews always divulge architectural details in layman's terms to hobby builders like myself, but such easily comprehensible info on why the optical flow thingy (is) was required for Nvidia's FG, and why it isn't for AMD's is awkwardly missing.

The TPU review of the 4090 for example said that the optical flow thingy is required for FG, that's why the feature is missing in Ampere and Turing. Then an article came sometime later stating that the OFA has always been part of Ampere as well as Turing, it just... ehm... (*nvidia PR person searching through notes*)... isn't fast enough.

It's propriety no amount of confirmation bias from Nvidia fans will objectively show us the basics on how this is now possible from the walled garden unless Jensen allows it. It has nothing to do with the backlash articles and public outcry from the CES and post CES marketing from Nvidia that they decided to throw a bone. Hey we have to give their PR team some praise it was clever of them to go with the eco friendly packaging on a 575 watt gpu after all for maximum virtue signal points. No critical thinking only someone talking down to you that you don't have a engineering degree to understand what's happening. Now the frame gen articles going to 3000 series are flooding the tech media and the backlash is yesterday's problem. Although this definitely raises more questions what the ef happens to the 4000 series flow accelerators. lol

 

[dyonoctis]

With all due respect, I shouldn't be required to read an 86-page long whitepaper full of programming jargon that, not being a programmer, I'll never understand, or to subscribe to an online journal just to know what a Tensor core is, or how it's connected to DLSS and FG. Reviews always divulge architectural details in layman's terms to hobby builders like myself, but such easily comprehensible info on why the optical flow thingy (is) was required for Nvidia's FG, and why it isn't for AMD's is awkwardly missing.

The TPU review of the 4090 for example said that the optical flow thingy is required for FG, that's why the feature is missing in Ampere and Turing. Then an article came sometime later stating that the OFA has always been part of Ampere as well as Turing, it just... ehm... (*nvidia PR person searching through notes*)... isn't fast enough.

It's a different approach to the problem. Nvidia has been public about the presence of optical flow hardware in Turing and ampere long before DLSS3 was a thing. AMD is also using optical flow for FSR 3, but it seems that they've managed to make it more efficient so that you don't need hardware acceleration. How? Nobody knows. it's all conjecture from observing the tech in action. AMD didn't explained. Now nvidia managed to make something that doesn't use optical flow at all, which is again different from what AMD did. But DLSS 3 and FSR are ironically closer than people thought . One is just more efficient at optical flow than the other.

That wouldn't be the first time that something like that happens in tech. Like how AMD can make smaller CPU cores that are better than Intel big CPU cores. Or how nvidia managed to develop some kind of compression algorhitm and tile rendering with maxwell that allowed that arch to do more with less. This time around AMD was the one who found a way to make optical flow more efficient.

AMD FSR 3 Frame Generation Analyzed

AMD's FSR 3 frame generation technology made a surprise debut earlier this month in two games, and we're ready to give you an early look, covering frame...

www.techspot.com

AMD FidelityFX™ Super Resolution 3 (FSR 3)

Discover frame generation with AMD FidelityFX™ Super Resolution 3, and get the source code and documentation!

gpuopen.com

FidelityFX Optical Flow 1.1.2 - FidelityFX SDK

FidelityFX Optical Flow 1.1.2 Introduction The FidelityFX Optical Flow technique estimates the motion between the current and the previous scene inputs. Based on AMD Fluid Motion Frames technology, and highly optimized for game inputs, the algorithm works on 8×8 blocks and calculates motion for...

gpuopen.com

Intel themselves is looking at another approach that won't introduce additional latency.

GFFE: G-buffer Free Frame Extrapolation for Low-latency Real-time...

Our method demonstrates how to generate extrapolated frames without adding input latency or extra buffers, simplifying integration into game engines to enhance user experience in real-time games.

www.intel.com

 

[Tomorrow]

It will be interesting to see whether RTX 3xxx series handle the optical flow.

Of course it will. If AMD was able to get it working while not using OFA then Nvidia surely could. It's a moot point as the new one uses Tensor cores, not OFA.

Yes, but... the VRAM capacity makes it a big problem for the RTX2000/3000 series at 1440p+ though, it could be beneficial on 1080p displays.

Yes it does increase VRAM usage but this mostly affects 6GB and 8GB cards. 11GB 2080Ti's and 12GB 3060's should be fine.

I mean, AMD is pretty much using the same kind of argument about FSR 4, so either everybody is lying, or there's some truth to what they are saying.

AMD has not ruled out bringing some version of FSR4 to older cards. Unlike Nvidia that flat out said it wont happen for 30 series.

Yes the Nvidia hate is strong. I'm not aware of a single feature they've ever artificially restricted to newer hardware,

ReBAR. RT. FG.

In any case I wouldn't get too excited for frame gen on 30 series, considering it may not improve performance as much as people expect on the older/slower hardware and it also uses up a bunch of VRAM that those cards don't really have to spare on modern games where frame gen would be most useful.

A few hundred MB extra. This only affects below 10GB cards the most. Also AMD already proved with FSR 3.1 that FG runs as well as DLSS FG on 30 series. I myself used it on 20 series with very few issues.

It's because you are listening to the vocal tiny minority that doesn't have access to them. Everyone else is busy enjoying dlss and FG.

What vocal minority? Thanks to AMD and LS everyone has access to AI upscaling and FG (even beyond Nvidia's 3x mode) on much older cards.

The same Nvidia fake promises, 2025 and not Resizable BAR for RTX2000 simply Nvidia don't want support for past generation GPUS the sold software updates like hardware upgrades

Exactly. ReBAR on 20 series would have been free performance, but required a new vBIOS that Nvidia did not bother making.
ReBAR is also not some super new AI tech that requires dedicated hardware. It has been in the PCIe spec for a long time.
AMD on the other hand supports ReBAR on their older cards too.

 

[Prima.Vera]

In any case I wouldn't get too excited for frame gen on 30 series, considering it may not improve performance as much as people expect on the older/slower hardware and it also uses up a bunch of VRAM that those cards don't really have to spare on modern games where frame gen would be most useful.

Bollocks!
AMD's FG works flawlessly on my RTX 3080, with minimum VRAM usage and latency, so you're telling me nVidia cannot do better? Hilarious.