NVIDIA would be working on AI-optimized drivers

Published by

Click here to post a comment for NVIDIA would be working on AI-optimized drivers on our message forum
https://forums.guru3d.com/data/avatars/m/258/258664.jpg
Ah yes, the magic drivers. If it exists, probably only for overpriced Lovelace.
https://forums.guru3d.com/data/avatars/m/296/296910.jpg
"driver is expected to be released in the first quarter of 2022" Just a typo, I think you must be referring to 2023.
https://forums.guru3d.com/data/avatars/m/235/235224.jpg
fantaskarsef:

Ah yes, the magic drivers. If it exists, probably only for overpriced Lovelace.
Doubt it, their drivers are universal fixes for their uarchs. Largest performance increase will always be latest arch given it's had the least amount of time for optimizations but the rest should benefit like previous enhancements except for that one time Kepler wasn't added to the improvements which was fixed with the next driver. I'm sure we'll be reminded about that soon enough through the echo-chamber.
https://forums.guru3d.com/data/avatars/m/258/258664.jpg
Spets:

Doubt it, their drivers are universal fixes for their uarchs. Largest performance increase will always be latest arch given it's had the least amount of time for optimizations but the rest should benefit like previous enhancements except for that one time Kepler wasn't added to the improvements which was fixed with the next driver. I'm sure we'll be reminded about that soon enough through the echo-chamber.
I hope you are right spets, but I am careful not to underestimate Nvidia's greed.
data/avatar/default/avatar02.webp
Most likely these are not drivers written by AI, but just optimized by AI. I can imagine that using AI to generate performance parameters for individual games and thus finding the best settings for performance is something that NVIDIA can easily do using their supercomputer clusters. Which GPU series will benefit is of course dependent on what NVIDIA deems economically viable. If I'd have to guess I would say that nothing below Turing will benefit from these AI optimized drivers.
https://forums.guru3d.com/data/avatars/m/258/258664.jpg
Something else: I guess it won't work without GFE? With the telemetry and all? 😀
https://forums.guru3d.com/data/avatars/m/268/268248.jpg
fantaskarsef:

Something else: I guess it won't work without GFE? With the telemetry and all? 😀
Either way I will still not install gfe 😛
https://forums.guru3d.com/data/avatars/m/275/275921.jpg
I try to keep my glass half full, so gonna go with "This could be good". Not gonna argue about GFE. Irrelevant to me if people use it or not.
https://forums.guru3d.com/data/avatars/m/258/258664.jpg
Venix:

Either way I will still not install gfe 😛
Honestly, it's everybody's choice, like @AuerX stated it too. I don't mind, as long as it offers something I like. Now, reassuring myself that the great days of recording my gameplay are no longer there, I don't need Shadowplay anymore, and thought about getting rid of GFE too (I normally install the vanilla driver). But reading this, first thought I had was "They probably make telemetry and GFE mandatory for an "AI driver", I bet!" Also, I think they could want to increase sales of older cards if they're the budget alternative, and get a boost (subjectively increasing the value of Ampere hardware), with such a driver. Makes me wonder, if they wouldn't be choking on old inventory, if such a miraculous driver would not have had to wait another year, or two, and artificially limited to newer hardware to boost e.g. Hopper adaption rates.
data/avatar/default/avatar32.webp
Optimizations that take better advantage of the int/FP units on Ampere/Ada? That's of course going to make everyone whine and cry about anything earlier not getting the same treatment (even though it wouldn't be possible).
https://forums.guru3d.com/data/avatars/m/259/259654.jpg
ttnuagmada:

Optimizations that take better advantage of the int/FP units on Ampere/Ada? That's of course going to make everyone whine and cry about anything earlier not getting the same treatment (even though it wouldn't be possible).
This would be amazing, as in theory there is a crazy amount of power in these cards.
https://forums.guru3d.com/data/avatars/m/63/63215.jpg
I'm sure I posted about using A.I. for the drivers before. I'm amazed it's not a thing until now.
https://forums.guru3d.com/data/avatars/m/198/198862.jpg
Up to 30%? Hopefully 30 series will have the driver not only 40 series.
https://forums.guru3d.com/data/avatars/m/242/242471.jpg
Main was avg 10% and up to 30%, like 4070ti is x3 faster then 3090ti in one game using dlss3 xD
https://forums.guru3d.com/data/avatars/m/251/251189.jpg
Undying:

Up to 30%? Hopefully 30 series will have the driver not only 40 series.
That'd be weird, as when not counting RT (& TCs?), afair the archs are extremely close.
data/avatar/default/avatar05.webp
Wording is the key here. "Up to 30%" Expect mostly less in most scenarios.
https://forums.guru3d.com/data/avatars/m/245/245459.jpg
This sounds like a good thing, like others have mentioned I'm expecting the performance improvements to be at least relevant for the 3000 series upwards (probably 2000 series upwards), although maybe the AI improvements don't have to be specific to certain architectures of GPUs. I reckon they'll offer the new drivers in the same way as before, letting you choose whether or not you install GFE, so still letting you get the driver seperately. Positive news!
data/avatar/default/avatar37.webp
I guess AMD could hire nvidia driver team to fix their garbage once nvidia switches to AI and fires them. 5000 series will cost even more because "AI drivers" is very expensive to print.
https://forums.guru3d.com/data/avatars/m/225/225084.jpg
If GFE is forced on us that dislike it then i'm sure if we kick off about it then Nvidia might not do it. Hopefully it's not the case.
https://forums.guru3d.com/data/avatars/m/246/246171.jpg
Seems like the kind of thing Intel really ought to be doing, considering they don't have the manpower to do optimizations. Nvidia has the most polished drivers already, so that explains why they're only going to see up to a 30% improvement, and I'm sure that won't be common.
ttnuagmada:

Optimizations that take better advantage of the int/FP units on Ampere/Ada? That's of course going to make everyone whine and cry about anything earlier not getting the same treatment (even though it wouldn't be possible).
As far as I understand, this is a pre-compiled optimization, not a realtime one. In other words, AI is used to figure out how to optimize the drivers for a game which are then hard-coded into the drivers.