Tensor Cores Useable For Other Things ?

[Dicehunter]

As we have a fair few Tensor cores inside the Turing line-up and going forward into next years lineup etc... which are a pretty cool technology for mixed precision training, Could these theoretically be used in games and programs for other things apart from DLSS ? It's a shame to see quite an advanced piece of tech not get used 95% of the time.

 

[Panos]

As we have a fair few Tensor cores inside the Turing line-up and going forward into next years lineup etc... which are a pretty cool technology for mixed precision training, Could these theoretically be used in games and programs for other things apart from DLSS ? It's a shame to see quite an advanced piece of tech not get used 95% of the time.

Doubt it for gaming, as they are too slow compared the normal shaders.

 

[Rroff]

Potentially they have a range of image manipulation and AI acceleration uses which is applicable to games but with the vendor specific nature and nVidia muddying the waters a bit as developers would be concerned with how much or little resources would be available if the user was using DLSS or other RTX features unlikely to be put into use any time soon.

Doubt it for gaming, as they are too slow compared the normal shaders.

Depends what you are doing - if the use of Tensor cores significantly accelerates what you are doing over other techniques that can be a net saving resulting in higher overall performance, if using Tensor cores doesn't reduce the time spent elsewhere (as in offloading to Tensor cores results in significantly reducing the time spent processing a feature via CPU resources or general shaders) as can be the case with DLSS in some situations then having the utilisation of them hanging on the end of a frame can result in a performance penalty overall - another thing that might make developers wary of using them in games if they can't easily foresee the impact.

 

[Panos]

Potentially they have a range of image manipulation and AI acceleration uses which is applicable to games but with the vendor specific nature and nVidia muddying the waters a bit as developers would be concerned with how much or little resources would be available if the user was using DLSS or other RTX features unlikely to be put into use any time soon.



Depends what you are doing - if the use of Tensor cores significantly accelerates what you are doing over other techniques that can be a net saving resulting in higher overall performance, if using Tensor cores doesn't reduce the time spent elsewhere (as in offloading to Tensor cores results in significantly reducing the time spent processing a feature via CPU resources or general shaders) as can be the case with DLSS in some situations then having the utilisation of them hanging on the end of a frame can result in a performance penalty overall - another thing that might make developers wary of using them in games if they can't easily foresee the impact.

I doubt game developers going to utilize the Tensor core for anything. Even DLSS requires the whole game been sent to Nvidia to provide the backed code and the executables.

 

[Rroff]

That won't be the case for everything using the Tensor cores though - DLSS requires a degree of offline learning to build reference patterns (ground truth) which won't be the case for other features using Tensor cores. Personally I think it is a horrid way to go about doing it while it has some academic merits and has lead to discovering approaches to better do other things but I'd much rather see the hardware estate used to provide higher quality AA without the performance hit you get currently while at native resolution.

 

[opethdisciple]

I could see something like the Adobe suite making use given developer support form both Adobe and Nvidia.