shrapnel wrote: ↑Mon Nov 11, 2019 2:45 pm
My understanding is that to be able to utilise cuddn-fp 16 you need to have an RTX 20 Series Card, like 2060, 2070, 2080 Super and of course the 2080 Ti Card.
He has an old GTX 1660, which will have only have cudnn as backend option,not cuddn-fp-16, unless of course he has the new GTX1660 Super,which he hasn't specified.
Even my previous 1080 Ti GTX which I had before I purchased my 2080 Tis refused to run if I chose cuddn-fp16 as Backend, I know this for my own experience. It ran only on plain cuddn.
You experts know perfectly well that Fat Fritz will run slow as molasses on an old GTX 1660, and he would feel cheated if he was buying Fritz 17 only for the NN part.
The whole GeForce 16 series is Turing based. I don't know what card you have the GTX 1660 confused with, but they were released this year.
My bad. If they are Turing based, then it will be able to utilize cuddn-fp16 as backend.
But still the performance will suffer and that's bad because lc0/fat fritz needs all the boost it can get, to be a Match for the top AB Engines.
My Card is: EVGA GeForce GTX 1660 SC ULTRA GAMING, 6GB GDDR5, Dual Fan, Metal Backplate, 06G-P4-1067-KR
It definitely states on box that it is "based on Turing Architecture".
I'm mostly interested in using FF as part of my analysis of my own miserable games, as well as position analysis in better ones; i.e. as an aid to my chess education.
Having watched LC0 a few times on TCEC I have been impressed by some moves but 'worried' by the frequent evaluation errors it exhibits.
In other words I'm used to waiting 10 mins for an insight into the position as well as checking the whole game automatically overnight using Chessbase 'Tactical Analysis'.
Sorry guys: you've been very helpful and I have no wish to take up more of your space/time here, though maybe others (equally objective!) might find this mini conversation of use.
Many thanks.
shrapnel wrote: ↑Mon Nov 11, 2019 2:45 pm
My understanding is that to be able to utilise cuddn-fp 16 you need to have an RTX 20 Series Card, like 2060, 2070, 2080 Super and of course the 2080 Ti Card.
He has an old GTX 1660, which will have only have cudnn as backend option,not cuddn-fp-16, unless of course he has the new GTX1660 Super,which he hasn't specified.
Even my previous 1080 Ti GTX which I had before I purchased my 2080 Tis refused to run if I chose cuddn-fp16 as Backend, I know this for my own experience. It ran only on plain cuddn.
You experts know perfectly well that Fat Fritz will run slow as molasses on an old GTX 1660, and he would feel cheated if he was buying Fritz 17 only for the NN part.
The whole GeForce 16 series is Turing based. I don't know what card you have the GTX 1660 confused with, but they were released this year.
My bad. If they are Turing based, then it will be able to utilize cuddn-fp16 as backend.
But still the performance will suffer and that's bad because lc0/fat fritz needs all the boost it can get, to be a Match for the top AB Engines.
The RTX 20xx series is Turing with TensorCores with fp16, int8... math,
the GTX 11xx series is Turing without TensorCores but with fp16 math support.
LC0 offers optimized backends for RTX series with TensorCores,
the TensorCores can speedup things 3x, or so, compared to non-TensorCores backends.
Not sure if LC0 has fp16 math backends for the GTX 11xx series,
afaik this was on Ankans todo list some months ago,
iirc he expected an 1.6 speedup or so compared to fp32 backend.
from LCZero releases:
v0.21.2-rc2
mooskagh released this on 22 May · 44 commits to master since this release
Changes
• Add 320 and 352 channel support for fused SE layer (#855)
• SE layer fix when not using fused kernel (#852) • Fp16 nchw for cudnn-fp16 backend (support GTX 16xx GPUs) (#849)
• Install lc0 on openSUSE - Full documentation, Install script and links to experimental RPM packages (#675)
LCZero is here 1.7 x faster with fp16 – and so will Fat Fritz do, I think.
jp wrote: ↑Mon Nov 11, 2019 5:41 am
But is the NN engine all you care about, not the Fritz17 AB engine? If that's the case, it's not worth buying.
Why not? For me seeing this result makes it interesting.
<snip>
Well, we don't have a huge number of examples so far in this BB, but it's been outperformed by Leela in analysis of positions, and the inquirer wants it to analyse his games.
Schwartmann wrote: ↑Mon Nov 11, 2019 7:30 pm
Will Fat Fritz be doing alright with GeForce GTX 1070 or will Stockfish be stronger on my Ryzen 7 2700x?
As in general rule, you get comparable strength of (20x256 net) Lc0 and Stockfish when they have 1:1000 nps speed ratio. GPU faster than that speed will crush SF and CPU faster than that speed will crush NN.
I highly doubt FatFritz will be stronger than Leela.
With your hardware setting Stockfish will be stronger than Fat Fritz or Leela.
Please forgive me if I missed it, but what is the size of the neural network of Fat Fritz? Ex: 10b, 20b, etc ...
Last edited by AdminX on Tue Nov 12, 2019 9:01 am, edited 1 time in total.
"Good decisions come from experience, and experience comes from bad decisions."
__________________________________________________________________
Ted Summers
