Fat Fritz Flummoxed

[Werewolf]

Fat Fritz on dual 2080 Ti (multiplexing, 4 threads, average nps = 55 Kn/s on this position) can't seem to find the right move after:

1.e4 Nf6
2.e5 Nd5
3.d4 d6
4.Nf3 de
5.Ne5 Nd7

and 6.Nf7 doesn't come even after 20 minutes thought...one of those quirks


[d]r1bqkb1r/pppnpppp/8/3nN3/3P4/8/PPP2PPP/RNBQKB1R w KQkq - 0 1[d]

Analysis by Fat Fritz (in lc0) rtx:

6.c4 N5f6 7.Nf3 e6 8.Nc3 c5 9.d5 Nb6 10.dxe6 Qxd1+ 11.Nxd1 Bxe6 12.b3 Be7 13.g3 a5 14.Bg2 a4 15.Rb1 Bf5 16.Rb2 axb3 17.axb3 Ra1 18.Bf4 0-0 19.0-0 Rfa8 20.Nh4 Bc8 21.Ne3 g6 22.Nf3 Rxf1+ 23.Kxf1 Ng4 24.Re2 Ra1+ 25.Re1 Rxe1+ 26.Kxe1 Nxe3 27.Bxe3 Nd7 28.Bf4 Nf8 29.Ng5 b6 30.Bd5 Ne6 31.Nxe6 fxe6 32.Bc6 Bd8 33.Ke2 Kf7 34.Kd3 Ke7 35.Ke4
White is slightly better: +/= (57% +39% =35% -26%) Depth: 25/72 00:12:31 34857kN
(Bicknell, 16.11.2019)

[shrapnel]

<img src="https://anildharan.tinytake.com/media/b ... height=655" title="Powered by TinyTake Screen Capture"/><br><a href="https://www.tinytake.com">Powered by TinyTake Screen Capture</a>
My lc0, using the Sergio NN, finds Nxf7 in 3-4 Seconds flat.
Odd what you are experiencing when you have a Setup similar to mine.
Maybe there is something wrong with Fat Fritz NN ?
BTW, why are you using 4 Threads ? I get maximum Kn/s using 3 Threads, any more, and it starts to dip.

[Nordlandia]

TCEC use 3 threads for 2x RTX cards.

Also why do you use not use fp16`?

[shrapnel]

TCEC use 3 threads for 2x RTX cards.

Also why do you use not use fp16`?

Are you talking to Werewolf or to me ?
Anyway, glad to know TCEC uses 3 Threads like me.
But Werewolf (and I) are correct to use "Multiplexing" in Backend and not just cuddn-fp16,as we are both running TWO 2080 Ti cards together, not just one !
This is what we put in Backend Options, cudnn-fp16(gpu=0),cudnn(gpu=1).
You would be correct to suggest fp-16 in Backend if we were running ONE 2080 Ti Card, which is not the case.

[Nordlandia]

TCEC use 3 threads for 2x RTX cards.

Also why do you use not use fp16`?

Are you talking to Werewolf or to me ?
Anyway, glad to know TCEC uses 3 Threads like me.
But Werewolf (and I) are correct to use "Multiplexing" in Backend and not just cuddn-fp16,as we are both running TWO 2080 Ti cards together, not just one !
This is what we put in Backend Options, cudnn-fp16(gpu=0),cudnn(gpu=1).
You would be correct to suggest fp-16 in Backend if we were running ONE 2080 Ti Card, which is not the case.

I was was Werewolf. Is it proven that Multiplexing is better for 2x 2080 Ti?

[shrapnel]

Is it proven that Multiplexing is better for 2x 2080 Ti?

Not only better, it is REQUIRED if you are using TWO 2080 Tis !
If you select cudnn-fp16 in backend, ONLY ONE, the Primary GPU will be utilized by lc0, causing a drastic reduction in Kn/s.
ONLY when Multiplexing( or demux or Round Robin) is enabled in backend, both GPUs are utilized by lc0.
You also have to have in backend options, backend=cudnn-fp16,(gpu=0),(gpu=1)
Hope it is clear now.

[Werewolf]

<img src="https://anildharan.tinytake.com/media/b ... height=655" title="Powered by TinyTake Screen Capture"/><br><a href="https://www.tinytake.com">Powered by TinyTake Screen Capture</a>
My lc0, using the Sergio NN, finds Nxf7 in 3-4 Seconds flat.
Odd what you are experiencing when you have a Setup similar to mine.
Maybe there is something wrong with Fat Fritz NN ?
BTW, why are you using 4 Threads ? I get maximum Kn/s using 3 Threads, any more, and it starts to dip.

I was advised by Albert to use either 3 threads with RoundRobin or 4 threads with Multiplexing. I tried both and Multiplexing was slightly faster.

I've also tried Lc0 here and it gets it very fast. Different weights = different results I suppose.

[Werewolf]

TCEC use 3 threads for 2x RTX cards.

Also why do you use not use fp16`?

Fp16 for ONE RTX card.

Multiplexing or RoundRobin for TWO

[Nordlandia]

Got it guys. Anyway Werewolf i recommend you to download "Real Temp" Real Temp is a temperature monitoring program.

Once both cards analyse you can open up the monitor tab and see how many percentages of your CPU is under load and adjust numbers of threads based on the listed percentage.

[shrapnel]

I was advised by Albert to use either 3 threads with RoundRobin or 4 threads with Multiplexing. I tried both and Multiplexing was slightly faster.

I've also tried Lc0 here and it gets it very fast. Different weights = different results I suppose.

Fair enough. And Yes, the weights definitely matter.
But I don't agree about the 4 Threads stuff, it definitely weakens lc0 if I put in 4 Threads and I don't suppose Fat Fritz is much different from lc0 from all I've read.
Why don't you put in 3 Threads this time and run your Test again and let us know if you get different results ?

Fp16 for ONE RTX card.

Multiplexing or RoundRobin for TWO

Short and sweet explanation.