Probably not enough tensor performance on Ampere to run the purely AI optical flow calculation faster than a hardware optical flow calculation. If it takes longer to use the AI approach, you don't see any benefit over the hardware approach.But I have a question for Jensen:
View attachment 114465
But it doesn't get the hardware approach either. And Intel can do it on a $250 B580 but the 3090 Ti cannot.Probably not enough tensor performance on Ampere to run the purely AI optical flow calculation faster than a hardware optical flow calculation. If it takes longer to use the AI approach, you don't see any benefit over the hardware approach.
🤷♂️ They're not using the OFA anymore. Does he still use Twitter I'll just go ask him and get him to say "not enough AI compute on a 3090 Ti" because I really do not believe it. It has all the TOPS.That's because Ampere's hardware optical flow accelerator is also too weak for frame gen.
Yes, it is predicting the future but it is not knowing the future.View attachment 114463
In all seriousness I think you nailed it. No matter how much leather tanning fumes someone might have huffed, no one believes NV is *telling the future* with DLSS4....
I get what you're saying, but in a purely forward predicting approach ("extrapolation") you are more subject to errors if there's sudden and random movement. Making the timesteps smaller just limits the amount of potential error between the last extrapolated frame and the "real" frame that comes after it, but the error can exist and it can be quite large if a lot of movement happens in that timestep. At least with interpolation, the start and end points are already known so you just need to fill in the gap, whether it be linear interpolation, polynomial, or whatever.Yes, it is predicting the future but it is not knowing the future.
Let me explain.
Imagine a baseball thrown horizontally at 90mph.
We are going to predict and compute it's path in two dimensions, x (horizontal) and y (vertical).
Computationally we have equations that can tell us how the ball will behave. Neglecting air resistance it's easy kinematics. With air resistance it's still possible but you need to use a differential equation and solve it numerically using something like 4th order Runge-Kutta.
For the prediction, we'll be stupid about it and simply compute a vector for the baseball's velocity (speed and direction) based on the previous two frames and move it ahead in time by the time of one frame.
The predicted location of the ball over such a short time interval will be pretty close to actual. In the horizontal direction it will only be off slighty due to air resistance. Vertically it will be off as it will assume constant vertical (downward) speed as opposed to acceleration due to gravity (-g).
It would be a simple matter to do the calculations and find the error but you will find it to be exceeding small even with our basically simple prediction method.
On the next computed frame the location of ball will be completely accurate.
Now if you remember the fundamental theory of Calculus then you will realize that as these time slices approact 0, the accuracy go to 100%. This is basically how finite element analysis works, you've got to have very good initial conditions, good equations to describe the conditions and environment, and exceedingly small time slices. It is numerically very expensive. As the time slices decrease the accuracy increases.
Have you ever wondered why over the last 50 or so years why weather "predictions" have become so much more accurate and detailed? It's due to the increase in computational power available coupled with the better initial condition data (weather balloons, satellite data, etc...).
I did a good bit of work in college as an engineer in college with FEA back in the stone age when we programmed in Fortran and ran the code on Prime computers. Things have moved on quite a bit but the underlying concepts are the same.
Of course now the prediction can be even better because the AI instantly recognize things and make appropriate actions like...
"Hey everything with mass looks to be moving in a gravitational field of strength g"
"That is a flash moving at light speed or some slowed down version to look good."
Those big creatures seem to move more slowly and have more inertia than the smaller ones."
So yeah this is all very possible and since not frames are interpolated there is no lag added to the game.
Probably not enough tensor performance on Ampere to run the purely AI optical flow calculation faster than a hardware optical flow calculation. If it takes longer to use the AI approach, you don't see any benefit over the hardware approach.
Yea, but a game that is well coded and balanced will NEVER scale anywhere near linear with increased resource of any kind, and that includes memory bandwidth. Even doing 0.5x scale is quite bad.It seems believable. A real shame if true, though. I still have hope that membw will do something in real games.
I think 4090 was itself not perfectly balanced. Lower on memory bandwidth than ideal, for example.Yea, but a game that is well coded and balanced will NEVER scale anywhere near linear with increased resource of any kind, and that includes memory bandwidth. Even doing 0.5x scale is quite bad.
Ideally the split is 1/3 shader, 1/3 fillrate, and 1/3 bandwidth, so if you double every resources, then you get double the performance. So in the ideal GPU/game scenario 2x bandwidth ~ 30% performance.
So 70% extra bandwidth is ~20% improvement overall, so it might be 5-10% extra over if the memory bandwidth was similar to the rest of the card, such as 30% more BW.
Because both the GPU and the games are an ever changing target.I think 4090 was itself not perfectly balanced. Lower on memory bandwidth than ideal, for example.
Oh. I was thinking Micron had failed them by about 2gbps.Because both the GPU and the games are an ever changing target.
They really aren't we've been in the era of deferred rendering with PBR mats for like a decade.Because both the GPU and the games are an ever changing target.
Paradigm shifts are trivial when you have a wooden shack. When you have built up that to a 100-storey skyscraper, radical changes are nearly impossible.In 7 years way back when we went from DX the original to DX9. You know a whole paradigm shift and stuff.
Guess what? It's been 7 years since DXR 1.0 too!
Well you don't. RTRT APIs like DXR are rudimentary DX7-era stuff.When you have built up that to a 100-storey skyscraper
They're very much possible but neither MS nor IHVs are trying to make RTRT FF thingies programmable.radical changes are nearly impossible.
Messed up formatting deleted. quotes got out of order. My apologies.What the heck man? This is just unrelated, uncalled for nonsense.
Also learn to read. I'm not praising NVidia anywhere here. I'm saying they should do better.
you know, Blackwell kinda feels like even NV gave up on pushing RTRT anywhere.They're very much possible but neither MS nor IHVs are trying to make RTRT FF thingies programmable.
What the heck man? This is just unrelated, uncalled for nonsense.
Also learn to read. I'm not praising NVidia anywhere here. I'm saying they should do better.
So either, accept i effed the quote formatting, and take it and leave it, or I will fix it tomorrow. sent you a DM, Your move.Messed up formatting deleted. quotes got out of order. My apologies.
EDIT: delete yours if you feel like it. Really the quoting system failed me.
You are correct but not really considering how small these time intervals are.I get what you're saying, but in a purely forward predicting approach ("extrapolation") you are more subject to errors if there's sudden and random movement. Making the timesteps smaller just limits the amount of potential error between the last extrapolated frame and the "real" frame that comes after it, but the error can exist and it can be quite large if a lot of movement happens in that timestep. At least with interpolation, the start and end points are already known so you just need to fill in the gap, whether it be linear interpolation, polynomial, or whatever.
In your analogy of predicting the flight of a baseball, that works well because a thrown baseball follows physics. What doesn't is user input, e.g. FPS games where I can be panning left, right, up, down or wherever and however I choose. It's that old stock investing adage of "past performance is no indicator of future results".