The paper includes the following chart for average frame gen times at various resolutions, in various test scenarios they compared with other frame generation methods.
Here’s their new method’s frame gen times, averaged across all their scenarios.
540p: 2.34ms
720p: 3.66ms
1080p: 6.62ms
Converted to FPS, by assuming constant frametimes, thats about…
540p: 427 FPS
720p: 273 FPS
1080p: 151 FPS
Now lets try extrapolated pixels per frametime to guesstimate an efficiency factor:
540p: 518400 px / 2.34 ms = 221538 px/ms
720p: 921600 px / 3.66 ms = 251803 px/ms
1080p: 2073600 px / 6.62 ms = 313233 px/ms
Plugging pixels vs efficiency factor into a graphing system and using power curve best fit estimation, you get these efficiency factors for non listed resolutions:
1440p: 361423 px/ms
2160p: 443899 px/ms
Which works out to roughly the following frame times:
1440p: 10.20 ms
2160p: 18.69 ms
Or in FPS:
1440p: 98 FPS
2160p: 53 FPS
… Now this is all extremely rough math, but the basic take away is that frame gen, even this faster and higher quality frame gen, which doesn’t introduce input lag in the way DLSS or FSR does, is only worth it if it can generate a frame faster than you could otherwise fully render it normally.
(I want to again stress here this is very rough math, but I am ironically forced to extrapolate performance at higher resolutions, as no such info exists in the paper.)
IE, if your rig is running 1080p at 240 FPS, 1440p at 120 FPS, or 4K at 60 FPS natively… this frame gen would be pointless.
I… guess if this could actually somehow be implemented at a driver level, as an upgrade to existing hardware, that would be good.
But … this is GPU tech.
Which, like DLSS, requires extensive AI training sets.
And is apparently proprietary to Intel… so it could only be rolled out on existing or new Intel GPUs (until or unless someone reverse engineers it for other GPUs) which basically everyone would have to buy new, as Intel only just started making GPUs.
Its not gonna somehow be a driver/chipset upgrade to existing Intel CPUs.
Basically this seems to be fundamental to Intel’s gambit to make its own new GPUs stand out. Build GPUs for less cost, with less hardware devoted to G Buffering, and use this frame gen method in lieu of that.
It all depends on the price to performance ratio.
Now this is all extremely rough math, but the basic take away is that frame gen, even this faster and higher quality frame gen, which doesn’t introduce input lag in the way DLSS or FSR does, is only worth it if it can generate a frame faster than you could otherwise fully render it normally.
The point of this method is that it takes less computations than going through the whole rendering pipeline, so it will always be able to render a frame faster than performing all the calculations unless we’re at extremes cases like very low resolution, very high fps, very slow GPU.
IE, if your rig is running 1080p at 240 FPS, 1440p at 120 FPS, or 4K at 60 FPS natively… this frame gen would be pointless.
Although you did mention these are only rough estimates, it is worth saying that these numbers are only relevant to this specific test and this specific GPU (RTX 4070 TI). Remember time to run a model is dependent on GPU performance, so a faster GPU will be able to run this model faster. I doubt you will ever run into a situation where you can go through the whole rendering pipeline before this model finishes running, except for the cases I listed above.
I… guess if this could actually somehow be implemented at a driver level, as an upgrade to existing hardware, that would be good
It can. This method only needs access to the frames, which can easily be accessed by the OS.
But … this is GPU tech.
This can run on whatever you want that can do math (CPU, NPU, GPU), they simply chose a GPU. Plus it is widely known that CPUs are not as good as GPUs at running models, so it would be useless to run this on a CPU.
And is apparently proprietary to Intel… so it could only be rolled out on existing or new Intel GPUs (until or unless someone reverse engineers it for other GPUs) which basically everyone would have to buy new, as Intel only just started making GPUs.
Where did you get this information? This is an academic paper in the public domain. You are not only allowed, but encouraged to reproduce and iterate on the method that is described in the paper. Also, the experiment didn’t even use Intel hardware, it was NVIDIA GPU and AMD CPU.
The point of this method is that it takes less computations than going through the whole rendering pipeline, so it will always be able to render a frame faster than performing all the calculations unless we’re at extremes cases like very low resolution, very high fps, very slow GPU.
I feel this is a bit of an overstatement, otherwise you’d only render the first frame of a game level and then just use this method to extrapolate every single subsequent frame.
Realistically, the model has to return back to actually fully pipeline rendered frames from time to time to re-reference itself, otherwise you’d quickly end up with a lot of hallucination/artefacts, kind of an AI version of a shitty video codec that morphs into nonsense when its only generating partial new frames based on detected change from the previous frame.
Its not clear at all, at least to me, in the paper alone, the average frequency, or under what conditions that reference frames are reffered back to… after watching the video as well, it seems they are running 24 second, 30 FPS scenes, and functionally doubling this to 60 FPS, by referring to some number of history frames to extrapolate half of the frames in the completed videos.
So, that would be a 1:1 ratio of extrapolated frame to reference frame.
This doesn’t appear to actually be working in a kind of real time, moderated tandem between real time pipeline rendering and frame extrapolation.
It seems to just be running already captured videos as input, and then rendering double FPS videos as output.
…But I could be wrong about that?
I would love it if I missed this in the paper and you could point out to me where they describe in detail how they balance the ratio of, or conditions in which a reference frame is actually referred to… all I’m seeing is basically ‘we look at the history buffer.’
Although you did mention these are only rough estimates, it is worth saying that these numbers are only relevant to this specific test and this specific GPU (RTX 4070 TI).
Thats a good point, I missed that, and it’s worth mentioning they ran this on a 4070ti.
I doubt you will ever run into a situation where you can go through the whole rendering pipeline before this model finishes running, except for the cases I listed above.
Unfortunately they don’t actually list any baseline for frametimes generated through the normal rendering pipeline, would have been nice to see that as a sort of ‘control’ column where all the scores for the various ‘visual difference/error from standard fully rendered frames’ are all 0 or 100 or whatever, then we could compare some numbers of how much quality you lose for faster frames, at least on a 4070ti.
If you control for a single given GPU then sure, other than edge cases, this method will almost always result in greater FPS for a slight degredstion in quality…
…but there’s almost no way this method is not proprietary, and thus your choice will be between price comparing GPUs with their differing rendering capabilities, not something like ‘do i turn MSAA to 4x or 16x’, available on basically any GPU.
More on that below.
This can run on whatever you want that can do math (CPU, NPU, GPU), they simply chose a GPU. Plus it is widely known that CPUs are not as good as GPUs at running models, so it would be useless to run this on a CPU.
Yes, this is why I said this is GPU tech, I did not figure that it needed to be stated that oh well ok yes technically you can run it locally on a CPU or NPU or APU but its only going to actually run well on something resbling a GPU.
I was aiming at practical upshot for average computer user not comprehensive breakdown for hardware/software developers and extreme enthusiasts.
Where did you get this information? This is an academic paper in the public domain. You are not only allowed, but encouraged to reproduce and iterate on the method that is described in the paper. Also, the experiment didn’t even use Intel hardware, it was NVIDIA GPU and AMD CPU.
To be fair, when I wrote it originally, I used ‘apparently’ as a qualifier, indicating lack of 100% certainty.
But uh, why did I assume this?
Because most of the names on the paper list the company they are employed by, there is no freely available source code, and just generally corporate funded research is always made proprietary unless explicitly indicated otherwise.
Much research done by Universities also ends up proprietary as well.
This paper only describes the actual method being used for frame gen in relatively broad strokes, the meat of the paper is devoted to analyzing it’s comparative utility, not thoroughly discussing and outlining exact opcodes or w/e.
Sure, you could try to implement this method based off of reading this paper, but that’s a far cry from ‘here’s our MIT liscensed alpha driver, go nuts.’
…And, now that you bring it up:
Intel filed what seem to me to be two different patent applications, almost 9 months before the paper we are discussing came out, with 2 out of 3 of the credited inventors on the patents also having their names on this paper, which are directly related to this academic publication.
This one appears to be focused on the machine learning / frame gen method, the software:
https://patents.justia.com/patent/20240311950
And this one appears to be focused on the physical design of a GPU, the hardware made to leverage the software.
https://patents.justia.com/patent/20240311951
So yeah, looks to me like Intel is certainly aiming at this being proprietary.
I suppose its technically possible they do not actually get these patents awardes to them, but I find that extremely unlikely.
EDIT: Also, lol video game journalism processional standards strike again, whoever wrote the article here could have looked this up and added this highly relevant ‘Intel is pursuing a patent on this technology’ information to their article in maybe a grand total of 15 to 30 extra minutes, but nah, too hard I guess.
I feel this is a bit of an overstatement, otherwise you’d only render the first frame of a game level and then just use this method to extrapolate every single subsequent frame.
Well, you would need a “history” of frames, so 1 wouldn’t be enough. Anyways, that’s fully possible, but then you would be generating garbage.
Realistically, the model has to return back to actually fully pipeline rendered frames from time to time to re-reference itself, otherwise you’d quickly end up with a lot of hallucination/artefacts, kind of an AI version of a shitty video codec that morphs into nonsense when its only generating partial new frames based on detected change from the previous frame.
That’s correct, nobody said otherwise. This is to help increase frame rate, so you need a source of frames to increase. Regular frames are still rendered as fast as the GPU can.
Its not clear at all, at least to me, in the paper alone, the average frequency, or under what conditions that reference frames are reffered back to… after watching the video as well, it seems they are running 24 second, 30 FPS scenes, and functionally doubling this to 60 FPS, by referring to some number of history frames to extrapolate half of the frames in the completed videos.
Because that’s implementation specific. As specified on the paper, once you have a history of frames, you can use the latest frame t_n to generate up to the t_(n+i) frame, where i is how many frames you want to generate. The higher i is, the higher the frame rate but also the more likely it is to be garbage.
So, that would be a 1:1 ratio of extrapolated frame to reference frame.
This doesn’t appear to actually be working in a kind of real time, moderated tandem between real time pipeline rendering and frame extrapolation.
I didn’t watch the video, but that’s completely possible. After you have a couple of frames generated, you can start alternating between a real frame and a generated one with this method. So you can’t have 60 fps at the beginning, but you can after a few frames.
It seems to just be running already captured videos as input, and then rendering double FPS videos as output.
The only difference between watching a movie and playing a video game is that the movie isn’t polling your input. This framework only cares about the previously rendered frames, and from a technical standpoint, they’re both just a bunch of pixels.
I would love it if I missed this in the paper and you could point out to me where they describe in detail how they balance the ratio of, or conditions in which a reference frame is actually referred to… all I’m seeing is basically ‘we look at the history buffer.’
Yes that’s because it is the implementer’s choice. I don’t know if they say what ratio they used, but it doesn’t matter because you don’t have to use their ratio. Anyone can implement this as they want and tune for quality/performance.
Unfortunately they don’t actually list any baseline for frametimes generated through the normal rendering pipeline, would have been nice to see that as a sort of ‘control’ column where all the scores for the various ‘visual difference/error from standard fully rendered frames’ are all 0 or 100 or whatever, then we could compare some numbers of how much quality you lose for faster frames, at least on a 4070ti.
Yes that’s a thing they seem to have missed. Would have been nice to see how it compared to actual rendering.
Yes, this is why I said this is GPU tech, I did not figure that it needed to be stated that oh well ok yes technically you can run it locally on a CPU or NPU or APU but its only going to actually run well on something resbling a GPU.
I was aiming at practical upshot for average computer user not comprehensive breakdown for hardware/software developers and extreme enthusiasts.
Yes that’s true for now. But remember that Windows started a trend with Copilot where manufacturers are now encouraged to include NPUs in their CPUs. Every modern laptop (M series, Qualcomm, latest Intel/AMD) now include NPUs in them (although underpowered ones, but these are first generation devices so it will inevitably get better), so in the near future these could run on the NPU that would come in almost all computers. Once NPUs are more common, this could easily become a driver.
To be fair, when I wrote it originally, I used ‘apparently’ as a qualifier, indicating lack of 100% certainty.
But uh, why did I assume this?
Because most of the names on the paper list the company they are employed by, there is no freely available source code, and just generally corporate funded research is always made proprietary unless explicitly indicated otherwise.
Much research done by Universities also ends up proprietary as well.
Yes Intel will not give the source code, but that’s not needed to recreate this experiment. Corporate funded academic research can be proprietary, but if it is published to the public then anyone is free to use that knowledge. The whole point of academic journals is to share the knowledge, if you wanted to keep it private you simply don’t publish it.
This paper only describes the actual method being used for frame gen in relatively broad strokes, the meat of the paper is devoted to analyzing it’s comparative utility, not thoroughly discussing and outlining exact opcodes or w/e.
Yes, because the method is all you need to recreate this. Intel is a for profit company so they might keep their own implementation to themselves. Pages 4:7 tell you exactly what you need to do to replicate this with details, they even give the formulas they used where needed. Remember this is supposed to be a general and modular framework that can be tuned depending on your goals, so the method needs to reflect that generality to allow for experimentation.
Sure, you could try to implement this method based off of reading this paper, but that’s a far cry from ‘here’s our MIT liscensed alpha driver, go nuts.’
They might publish it in the future, they might not, but if they don’t nothing is lost and they get a head start on implementing research that they paid for.
Intel filed what seem to me to be two different patent applications, almost 9 months before the paper we are discussing came out, with 2 out of 3 of the credited inventors on the patents also having their names on this paper, which are directly related to this academic publication.
This one appears to be focused on the machine learning / frame gen method, the software:
https://patents.justia.com/patent/20240311950
This patent is about hardware configuration of a system designed to run such a model in a way that Intel considers optimal. So I guess they’re considering designing SOCs specialized on these things (maybe for handhelds?). But this is not related to the paper, since this doesn’t affect your ability to train and run this model on your RTX like they did on the paper.
And this one appears to be focused on the physical design of a GPU, the hardware made to leverage the software.
https://patents.justia.com/patent/20240311951
So yeah, looks to me like Intel is certainly aiming at this being proprietary.
I suppose its technically possible they do not actually get these patents award
This one is more tricky, but it also does not affect your ability to implement your own model. What they are doing here is akin to a real-time kernel operation but for graphics. You set a maximum time for a frame to be rendered (ideally monitor refresh rate), if the algorithm decides that the GPU won’t meet that deadline, then you generate the frame and discard whatever the GPU was doing. It’s basically a guarantee to meet the display update frequency (or proper v-sync). Also they aren’t likely to get this one because they’re trying to patent the logic: if time1 is less than tmax, pick option one; else pick option two.
These patents do not affect the paper in any way, since they do not cover what is needed for this method (RTX 4070 Ti, Ryzen 9 5900X, Pytorch, TensorRT, and NVIDIA Falcor) or their alternatives.
