@DegustatoR I'm guessing the SM changes probably have something to do with scheduling ai work and interleaving neural shaders with other shaders. Not sure. That's seemed to be what was being suggested when Jensen said something like, "We can do them at the same time."
Man from Atlantis
Veteran
According to Richard at DF, MFG is limited to 50 series because of framepacing hardware (timestamped) on Blackwell GPUs.
DegustatoR
Legend
Yeah, well, they could've done shading and tensor operations "at the same time" since Ampere. It is usually impossible due to register file and memory bandwidth though, not because of how the SM is designed. So it is unclear what has changed in Blackwell. Maybe they've moved the tensor ALUs inside main shading SIMDs and they all are controlled by the same logic now? Would be wild as that's seem to be how AMD has implemented AI h/w in RDNA4, and also would actually be a step backwards from how tensor h/w was built into Nvidia GPUs since Volta.
So if framegen is done in Tensor cores now could/would they back port it to Ampere or even Turing? Not multi frame but single?
I am assuming no because I feel from a business perspective trying to explain why it’s possible now but wasn’t possible when the 40 series came out will be difficult unless the audience is technically knowledgeable.
I am assuming no because I feel from a business perspective trying to explain why it’s possible now but wasn’t possible when the 40 series came out will be difficult unless the audience is technically knowledgeable.
Man from Atlantis
Veteran
I mean, they could say something along these lines that the software and AI training model wasn’t ready back then, similar to how DLSS evolved into DLSS 2, and they beefed up the existing OFA hardware to handle the extra workload for the time being. Now that it's ready, 3080/3090 people can enjoy 2x FG
I would certainly appreciate it if they did this. They are deploying the new DLSS models to as far back as Turing so frankly at this point despite the high prices id argue Nvidia is the most consumer friendly in the business right now.I mean, they could say something along these lines that the software and AI training model wasn’t ready back then, similar to how DLSS evolved into DLSS 2, and they beefed up the existing OFA hardware to handle the extra workload for the time being. Now that it's ready, 3080/3090 people can enjoy 2x FG
One thing I've not seen discussed is whether frame gen can actually improve the overall frame pacing of a game. Presumably it can since it will know the frame delivery time of the two real frames and then pace the 3 generated frames accordingly to smooth out that average.
That might be BS, but I have definitely noticed games feeling less juddery with FG on which can't be accounted for by the higher number of frames alone. In Indiana Jones for example, regardless of what I did with the settings to get a higher frame rates, the game just didn't feel smooth to me. But when I turned on FG it felt like butter, even after ramping up settings to the point where the end FG frame rate wasn't that much higher than the native frame rate I was using previously which felt much less smooth.
That might be BS, but I have definitely noticed games feeling less juddery with FG on which can't be accounted for by the higher number of frames alone. In Indiana Jones for example, regardless of what I did with the settings to get a higher frame rates, the game just didn't feel smooth to me. But when I turned on FG it felt like butter, even after ramping up settings to the point where the end FG frame rate wasn't that much higher than the native frame rate I was using previously which felt much less smooth.
I think so, TVs will smooth out 3:2 pull down judder with frame insertion.
Ext3h
Regular
Sounds just like there might be another set of tensor ops, that are better suited for loading a (potentially smaller) tensor from only 4 threads of a warp each. There used to be the gotcha that 16x16/8x32/32x8 was the smallest supported tensor size (for any argument) and the native size of the tensor cores kept creeping up towards that.
Plus, so far the whole warp needs to be cooperative, so currently that would be 8 rounds till all 8 quads in a warp got their turn to use the dedicated set of registers that the tensor core could access on a SM. So most likely the number of registers accessible from the tensor cores has been ramped up.
What you’re getting is better motion performance from your monitor as the frame rate gets closer to the monitor refresh rate which increases motion clarity, perceived and actual.
It’s exactly what FG is meant to do.
I’m upgrading to a 240Hz OLED monitor next week and looking forward to seeing what shorter hold times feel like. Not playing any frame generation games anytime soon but hopefully I can get older games up near the monitor limit.
DegustatoR
Legend
It depends. You can already see improvements in some titles with DLSS FG on 40 series but there can only be improvements when "real frames" frametimes are relatively stable. If you're getting hitching - and if it's also due to some other bottleneck intervening into rendering like a sudden CPU limitation for one frame in a sequence of GPU limited frames - then FG actually make them worse. The latter is why some people are saying that FG adds stuttering sometimes - it's not FG per se but it's the stutters and hitches of the base game which are getting amplified by FG to a degree.
While we talk about games, just the smoothness of the os, moving windows, scrolling, mouse etc will feel so fluid. Enjoy it,
Other than MFG, will all the other stuff be available on older cards? I think so but there's too much information for me to sort through.
This is better pricing than I expected across the board. I'm especially surprised by the 5090. Was expecting at least $2500 for that monster.
This is better pricing than I expected across the board. I'm especially surprised by the 5090. Was expecting at least $2500 for that monster.
Just watch the DF video. Everything relevant there.
DavidGraham
Veteran
I hope we can all learn from this launch, all the prices listed before launch were place holders or fakes, this happens every launch, the only one with knowledge of prices is NVIDIA on launch day.
DegustatoR
Legend
In DLSS suite yes.
In their RTX Kit though I suspect not all tech will work on all cards. Things like RTX Hair (not a good name btw) seem to use new features of Blackwell RT cores (new sphere primitive type). Some other stuff is unclear if it will work well on previous gen h/w (neural shaders for example).
DavidGraham
Veteran