The optimalized Fat Fritz

[corres]

Albert Silver wrote an article about optimization of Fat Fritz.
https://en.chessbase.com/post/optimizing-fat-fritz
I made some test to investigate the effect of the optimization on Fat Fritz.
Machine:
AMD Ryzen Threadripper 2950x 16 x 4 GHz, HT/SMT = OFF, turbo = OFF, no throttling
2 x RTX 2060 OC
RAM 32 GB
OS: Windows 10 64 bits
Chess GUI: DeepFritz 14

NN engines:
Lc0 ver.0.22.0
Fat Fritz on lc0 ver.1.0
Parameters of engines:
Lc0 parameters
-Threads 4
-Backend multiplexing
-BackendOptions backend=cudnn-fp16,(gpu=0),(gpu=1)
-NNCacheSize 10000000
-MaxPrefetchBatch 64
Others are Default.
Fat Fritz parameters
a, Default Fat Fritz (Fat Fritz-def) parameters excluding
-Threads 4
-Backend multiplexing
-BackendOptions backend=cudnn-fp16
-NNCacheSize 10000000
Others are Default
b,Using Lc0 parameters (Fat Fritz-lc0) - every parameters of Fat Fritz are modified to Lc0 parameters.

Test params:
-Time Control 1 min + 2 sec/ move
-Openings 50 position (my TestBook50) with alternated colors

Results
1, Fat Fritz against 42872 net
a, 42872 : Fat Fritz-def = 6 : 3 (91 draw) 100 games
b, Fat Fritz-lc0 : 42872 = 6 : 3 (91 draw) 100 games
It seems the optimized Fat Fritz becomes weaker against 42872 net.

2, Fat Fritz against T40B.4-160 net
a, T40B.4-160 : Fat Fritz-def = 9 : 5 (86 draw) 100 games
b, T40B.4-160 : Fat Fritz-lc0 = 12 : 5 (83 draw) 100 games
In this case the optimization enhanced the power of Fat Fritz in some measure (+ 3 games).
But the optimized Fat Fritz with its own net is also weaker then Lc0 with T40B.4-160.

3, Fat Fritz against Stockfish 191002
a, Stockfish191002 : Fat Fritz-def = 17 : 11 (72 draw) 100 games
b, Stockfish191002 : Fat Fritz-lc0 = 20 : 8 (72 draw) 100 games
Against Stockfish the optimization also overhauled Fat Fritz (+ 6 games)
But Fat Fritz on 2 x RTX 2060 OC (~ RTX 2080 Ti OC) is obviously weaker than Stockfish on a 16 x 4 GHz PC.

[corres]

Albert Silver wrote an article about optimization of Fat Fritz.
https://en.chessbase.com/post/optimizing-fat-fritz
I made some test to investigate the effect of the optimization on Fat Fritz.
Machine:
AMD Ryzen Threadripper 2950x 16 x 4 GHz, HT/SMT = OFF, turbo = OFF, no throttling
2 x RTX 2060 OC
RAM 32 GB
OS: Windows 10 64 bits
Chess GUI: DeepFritz 14

NN engines:
Lc0 ver.0.22.0
Fat Fritz on lc0 ver.1.0
Parameters of engines:
Lc0 parameters
-Threads 4
-Backend multiplexing
-BackendOptions backend=cudnn-fp16,(gpu=0),(gpu=1)
-NNCacheSize 10000000
-MaxPrefetchBatch 64
Others are Default.
Fat Fritz parameters
a, Default Fat Fritz (Fat Fritz-def) parameters excluding
-Threads 4
-Backend multiplexing
-BackendOptions backend=cudnn-fp16,(gpu=0),(gpu=1) <<-- CORRECTION!
-NNCacheSize 10000000
Others are Default
b,Using Lc0 parameters (Fat Fritz-lc0) - every parameters of Fat Fritz are modified to Lc0 parameters.

Test params:
-Time Control 1 min + 2 sec/ move
-Openings 50 position (my TestBook50) with alternated colors

Results
1, Fat Fritz against 42872 net
a, 42872 : Fat Fritz-def = 6 : 3 (91 draw) 100 games
b, Fat Fritz-lc0 : 42872 = 6 : 3 (91 draw) 100 games
It seems the optimized Fat Fritz becomes weaker against 42872 net.

2, Fat Fritz against T40B.4-160 net
a, T40B.4-160 : Fat Fritz-def = 9 : 5 (86 draw) 100 games
b, T40B.4-160 : Fat Fritz-lc0 = 12 : 5 (83 draw) 100 games
In this case the optimization enhanced the power of Fat Fritz in some measure (+ 3 games).
But the optimized Fat Fritz with its own net is also weaker then Lc0 with T40B.4-160.

3, Fat Fritz against Stockfish 191002
a, Stockfish191002 : Fat Fritz-def = 17 : 11 (72 draw) 100 games
b, Stockfish191002 : Fat Fritz-lc0 = 20 : 8 (72 draw) 100 games
Against Stockfish the optimization also overhauled Fat Fritz (+ 6 games)
But Fat Fritz on 2 x RTX 2060 OC (~ RTX 2080 Ti OC) is obviously weaker than Stockfish on a 16 x 4 GHz PC.

[shrapnel]

So what's the Conclusion ?

Fat Fritz SUX ?

[Nordlandia]

Fat Fritz is said to get new weight_file soon.

Anyway do you guys urge me to buy another RTX 2080 Super for 2x 2080 Super OC in NVlink.

2x 2080 Super OC (Asus) is said be about equivalent to one and a half 2080 Ti in performance.

[corres]

So what's the Conclusion ?
Fat Fritz SUX ?

Fat Fritz is further developed and better version of DeusX.
As every engines Fat Fritz has the own weak points and strong points.
Only the practical user - like you - can decide about Fat Fritz.
I used only short time control and a medium powered hardware with relative few games.
So the final conclusion should be said by the community of chess engine users.
Note
If I were you I would try the parameters of Fat Fritz on Leela with J13B.2-188 net.
Parameters of Fat Fritz give more aggressive playing than the default Leela parameters.
Maybe Leela will be stronger with them.

[corres]

Fat Fritz is said to get new weight_file soon.
Anyway do you guys urge me to buy another RTX 2080 Super for 2x 2080 Super OC in NVlink.
2x 2080 Super OC (Asus) is said be about equivalent to one and a half 2080 Ti in performance.

It depends on the used net and the used RTX product but 2 x RTX 2080 Super OC ~ 1.7 x RTX 2080 Ti in chess power.

[shrapnel]

So what's the Conclusion ?
Fat Fritz SUX ?

Fat Fritz is further developed and better version of DeusX.
As every engines Fat Fritz has the own weak points and strong points.
Only the practical user - like you - can decide about Fat Fritz.
I used only short time control and a medium powered hardware with relative few games.
So the final conclusion should be said by the community of chess engine users.
Note
If I were you I would try the parameters of Fat Fritz on Leela with J13B.2-188 net.
Parameters of Fat Fritz give more aggressive playing than the default Leela parameters.
Maybe Leela will be stronger with them.

Tell me the parameters, I will check it out.

[corres]

So what's the Conclusion ?
Fat Fritz SUX ?

Fat Fritz is further developed and better version of DeusX.
As every engines Fat Fritz has the own weak points and strong points.
Only the practical user - like you - can decide about Fat Fritz.
I used only short time control and a medium powered hardware with relative few games.
So the final conclusion should be said by the community of chess engine users.
Note
If I were you I would try the parameters of Fat Fritz on Leela with J13B.2-188 net.
Parameters of Fat Fritz give more aggressive playing than the default Leela parameters.
Maybe Leela will be stronger with them.

Tell me the parameters, I will check it out.

I do not know how like it Albert Silver.
Maybe they are patented.
Please, ask him.

[corres]

So what's the Conclusion ?
Fat Fritz SUX ?

Fat Fritz is further developed and better version of DeusX.
As every engines Fat Fritz has the own weak points and strong points.
Only the practical user - like you - can decide about Fat Fritz.
I used only short time control and a medium powered hardware with relative few games.
So the final conclusion should be said by the community of chess engine users.
Note
If I were you I would try the parameters of Fat Fritz on Leela with J13B.2-188 net.
Parameters of Fat Fritz give more aggressive playing than the default Leela parameters.
Maybe Leela will be stronger with them.

Tell me the parameters, I will check it out.

I do not know how like it Albert Silver.
Maybe they are patented.
Please, ask him.

Note
In the cited article of Albert Silver you can see parameters of Fat Fritz on pictures.
Try them.

[corres]

You can download the test files:
http://wikisend.com
Download
File ID 322386
Password fatfritztest