RTX 2080 Ti (and other 20xx) revealed

[frakty]

Anyone followed the new nvidia series cards reveal?

Assuming information on number of tensor cores here is correct

https://www.pcper.com/news/Graphics-Cards/NVIDIA-Announces-GeForce-RTX-2070-RTX-2080-and-RTX-2080-Ti-Gamescom-2018

this could be potentially quite significant for finding best hardware for lc0 value / money performance. 

I am still a bit reserved about it because when revealing the new Quadro series they were much more explicit about the tensor cores (giving exact flops numbers etc.), while here it looked like they were a bit overlooking it.

So at the moment the only thing that's confirmed is that 2080 Ti tensor cores can be used for AI image post-processing like for example ray traced scene de-noising or temporal anti-aliasing. I am wondering if they could be used also outside of image post-processing (if they are fully unlocked for external applications like e.g. lc0) 

Any insights on this? If leela could really take advantage of those tensor cores am I correct to assume that would be awesome?

[TheHomicidalTendency]

Watched the reveal live, if I remember correctly the tensor cores in the 2080ti at least are 110 tflop   performance. I was more worried if Lchess was even setup to allow them to be used. But, you're right Nvidia probably doesn't want consumer based gpu's being able to be used for machine learning algorithms where they sell their $10,000 + cards. That being said does this program allow tensor cores? If so how would it compare to google's program or a 1080 ti? (new here)

[frakty]

Here is some spreadsheet with nps on various cards, including Titan V and GTX 1080 Ti

https://docs.google.com/spreadsheets/d/1lGFf6PLGmBUSMan-YP7Vul4DpRNfn6K8oeCjBILe6uA/edit#gid=1508569046

While Titan V also provides tensor cores and 110  TFLOPs for deep learning (theoretical) in practice it seems most of it's speed up compared to GTX 1080 Ti comes from better support of fp16 precission. Still the ~ 3x speedup relative to 1080 Ti is significant.

Perhaps someone more knowledgeable could explain if with implementation better optimized for tensor cores the speedup on cards like Titan V or V100 could be even greater than it currently is.

[frakty]

There is a rumor posted on talkchess.com that someone who works for nvidia driver development (and who has interest in chess engines) posted on discord that: 

"Rtx 2080ti is almost same as titan v for lc0"

I still don't take it as 100% confirmation of course but I would recommend that anyone who considers buying GPU for leela should wait until Sep 20th (when the new cards become available) as that could potentially result in much improved speed / money ratio.