Sergio NN 40x512

[Nordlandia]

It's early in training. Too early to draw conclusions but it perform subpar.

[corres]

...
Yes I don't know why the testing sites only play super lightning blitz tournament on only default settings and not the best settings except for me to rate NN engines. I am glad I do not. I guess they stupidly assume that all NN are created equal in regards to the default settings.

If you have a lot of time to waste please, make a line of tests to optimize the parameters and the another line of tests to show the relations among nets.

I have found the bigger nets to be better at shorter time controls and longer time controls. NPS is a very poor gauge on what is the best NN at any time control. I have show this in my testing.
...

Where are the arguments?
The speed of a net is such an important parameter than the knowledge stored in NN. A "clever" but too slow net is not better then a fast but "stupid" net. The task of net-makers is for making nets with optimal speed and knowledge.

Lets see where Sergio 40x512 stands in regards to the other engines and NN.

[mwyoung]

...
Yes I don't know why the testing sites only play super lightning blitz tournament on only default settings and not the best settings except for me to rate NN engines. I am glad I do not. I guess they stupidly assume that all NN are created equal in regards to the default settings.

If you have a lot of time to waste please, make a line of tests to optimize the parameters and the another line of tests to show the relations among nets.

I have found the bigger nets to be better at shorter time controls and longer time controls. NPS is a very poor gauge on what is the best NN at any time control. I have show this in my testing.
...

Where are the arguments?
The speed of a net is such an important parameter than the knowledge stored in NN. A "clever" but too slow net is not better then a fast but "stupid" net. The task of net-makers is for making nets with optimal speed and knowledge.

Lets see where Sergio 40x512 stands in regards to the other engines and NN.

Yes testing with correct settings. That is to stupid. I guess we know you are the problem. To bad I am here to show you are full of shit.
Testing with correct settings is self evident.

[shrapnel]

The speed of a net is such an important parameter than the knowledge stored in NN. A "clever" but too slow net is not better then a fast but "stupid" net. The task of net-makers is for making nets with optimal speed and knowledge.

I would say that it is rather YOUR Task, not NN maker's, to make sure you have the fastest possible GPU to fully UTILIZE the large NN.
Its rather silly to use a slow Graphic Card for Testing and then say that large NN is bad !
You neglect the importance of Hardware for testing NNs.

[Hai]

A 512x40 test is running here:

https://www.comp.nus.edu.sg/~sergio-v/test/

Thank you very much. It looks much better running LC0 Sergio 40x512 net with my 2x RTX 2080 Ti.

This is amazing

Hello Hai. What speed do you get in this FEN position. I'm asking because i want to compare the speed difference. I'm running 2x RTX 2080 Super OC (ASUS ROG STRIX) Identical cards.

[d]1q6/1r2k1p1/4pp1p/1P1b1P2/3Q4/7P/4B1P1/2R3K1 w

Analyse for 1 minute with 512x40 net.

I'm using following settings tweaked from the defaults:

<------->
Roundrobin
(backend=cudnn-fp16,gpu=0),(backend=cudnn-fp16,gpu=1)
CPuct: 3.4
CPuctBase: 10000
MiniBatchSize: 512
<------->

Tablebases disabled.

Assuming you have two identical cards try same settings or multiplexing if they're not intentical and run that test for speed comparison

1 minute search:

1q6/1r2k1p1/4pp1p/1P1b1P2/3Q4/7P/4B1P1/2R3K1 w - - 0 1

Analysis by Lc0 v0.23.2+git.c8d9095:

1.Qg4 exf5 2.Qxg7+ Ke6 3.Qg6 Qe5 4.Qe8+ Re7 5.Qg8+ Rf7 6.Re1 Kd6 7.b6 Rd7 8.Qf8+ Re7 9.Qb8+ Ke6 10.Qc8+ Kd6 11.Kh1 Bxg2+ 12.Kxg2 Qe4+ 13.Kf2 Qe3+ 14.Kf1
White is better: +/- (1.14) Depth: 12/35 00:01:06 329kN
(12.01.2020)

2x 2080 Super output: speed 5544 N/s after about one minute.

Hi Nordlandia,

we have nearly the same N/s and kN.

I'm using default settings.
And also:
Roundrobin
(backend=cudnn-fp16,gpu=0),(backend=cudnn-fp16,gpu=1)
+ 64 GB RAM

You have 2x 2080 Super and OC.
I have 2x 2080 Ti without Super and without OC.
Also I use power limit of 95% per GPU.
Also I use temperature limit of 70 per GPU.
So I've got lower power consumption, lower heat and it's also really quiet .
But GPU load is jumping between 60 to 90% on both GPUs when running your position.

When I run chess start position I have GPU load of 93% on both GPUs.
And CPU heat is 50.

I guess the question is does this network 512x40 need 1, 2,3,4....High end graphic cards. Assuming it is a good net to be the best. Have you run any game test with this Fat Network. I will download it for a full round of testing. Have we now reached the limits of 1 RTX graphics card?

I can test if this Fat Network can top the best on 1 RTX....

I'm interested in chess analysis and not in engine-super-lightning-bullet-tc-chess .

Like for every big net, at the beginning you should have 2x RTX GPUs, 2080 Ti if possible.
Like for every big net, later one RTX GPU, 2080 Ti if possible, will be enough.
-30x384-T60-2, 131 MB, : Ongoing, at 3rd LR (0.002) - That explains why 1 RTX GPU is enough.
-40x512, 318 MB, already started to train = less knowledge at the moment = to compensate it you sould use 2x RTX GPUs. But it also benefit much more from 3 or 4 RTX GPUs than the other nets.

-I can do now much longer analysis, when using the LC0 Sergio 40x512 net.
-I need much lesser RAM when running the 40x512 net.
-It is also possible to run it much longer than a 30x384 net before I run out of memory with my 64 GB RAM.

I expect the 40x512 net to get ~60 points in the ERET Test, because it's an untrained net at the moment.
Of course when using longer time control, the result will be better.
I have tested it also for analysis and games playing style.
It can already find some very difficult moves when using it on some fen positions to analyse. That's really amazing.
Strategically it is weaker at the moment and the positional style is also weaker at the moment.
But tactics are better .

Yes I don't know why the testing sites only play super lightning blitz tournament on only default settings and not the best settings except for me to rate NN engines. I am glad I do not. I guess they stupidly assume that all NN are created equal in regards to the default settings.
I have found the bigger nets to be better at shorter time controls and longer time controls. NPS is a very poor gauge on what is the best NN at any time control. I have show this in my testing.

And I assume you know that every position in a chess game is a series of positions to analyze. And the engine that is most correct general wins the game. That is why we test. General the strongest chess playing engine will give you the best analysis.

Lets see where Sergio 40x512 stands in regards to the other engines and NN.

I have very very very good news .
I've tested the frist 40x512 net with the ERET test suite.
First result with 8GB RAM and a limit of 15 seconds per move: 59 of 111 points.
Second result with 64 GB RAM and a limit of 300 seconds per move: 82 of 111 points.

This is an improvement of 23 points holy shit
83 / 111 = 74.7%.
Mean time = 36.40s / 9.97
Usually when doing the same test as described before with the smaller nets, the improvement was only ~2 points more than before.
This is the first net which have the capability to solve the complete ERET Test Suite after it is trained.

[peter]

I have very very very good news .
I've tested the frist 40x512 net with the ERET test suite.
First result with 8GB RAM and a limit of 15 seconds per move: 59 of 111 points.
Second result with 64 GB RAM and a limit of 300 seconds per move: 82 of 111 points.

This is an improvement of 23 points holy shit
83 / 111 = 74.7%.
Mean time = 36.40s / 9.97
Usually when doing the same test as described before with the smaller nets, the improvement was only ~2 points more than before.
This is the first net which have the capability to solve the complete ERET Test Suite after it is trained.

First of all, you don't do any "training" of a NN by simply letting it play the same positions twice in playing or analysing mode. You have to let it play in learning mode, you didn't do so, did you?
So where's your point about "capability to solve the complete ERET test suite" after the big net was trained at certain positions?

I'd be rather sure it would be much easier to train a smaller net especially for certain single positions than a bigger one, much less games would be to be played for that purpose. To be sure how many for which result of which net at what a test- TC to be measured, you'd have to try just that.

Second, of course you have much more solutions with 300" per move compared to 15" at the first try, 15"/position is, what Walter Eigenmann recommends, 59 /111 with this TC is rather weak and 83 at 300" isn't a really good result neither. Much smaller nets do that at good hardware since quite a while in 15"/position too.

Third, if you want to see the learning effect of an engine at Eret, look here:

Code: Select all

Houdini 4 Learn, 109/111
Löse: C:\...\ERET - CBH\Eigenmann Rapid Engine Test.cbh
Maximale Lösungszeit = 15s.

1. ERET 001 - Entlastung,  Akopian-Khenkin, Tilburg 1994   Gelöst in 0.02s/7; Gelöst: 1
2. ERET 002 - Zugzwang,  Akerblom 1961   Gelöst in 0s/7; Gelöst: 2
3. ERET 003 - Linienoeffnen,  Alvarez-Kokkila, CorrGame 1997   Gelöst in 0s/7; Gelöst: 3
4. ERET 004 - Endspiel L&S vs L&S,  Anand-Shirov, Advanced Chess 2000   Gelöst in 0s/7; Gelöst: 4
5. ERET 005 - Laeuferopfer f7,  Andriulaitis-Ruefenacht, CorrGame 1999   Gelöst in 0s/7; Gelöst: 5
6. ERET 006 - Springeropfer,  Annageldyev-Akopian, Moskau 1994   Gelöst in 0s/7; Gelöst: 6
7. ERET 007 - Laeuferpaar,  Aronian-Anand, WiikAanZee 2013   Gelöst in 0s/7; Gelöst: 7
8. ERET 008 - Zentrum,  Aronian-Kramnik, Monte Carlo 2011   Gelöst in 0.05s/7; Gelöst: 8
9. ERET 009 - Springeropfer,  Aseev-Popv, Kazan 1995   Gelöst in 0.02s/7; Gelöst: 9
10. ERET 010 - Freibauer,  Atalik-Bregadze, St Louis 2013   Gelöst in 0s/7; Gelöst: 10
11. ERET 011 - Rochadeangriff,  Atalik-Sax, Szeged 1997   Gelöst in 0s/7; Gelöst: 11
12. ERET 012 - Entlastung,  Babula-Hoffmann, Dt-BL 2011   Gelöst in 0.05s/8; Gelöst: 12
13. ERET 013 - Zentrum,  Bang-Umansky, CorrGame 1996   Gelöst in 0.91s/17; Gelöst: 13
14. ERET 014 - Endspiel T&L&S vs T&L&S,  Barbre-D'Arruda, CorrGame 1994   Gelöst in 0.02s/9; Gelöst: 14
15. ERET 015 - Endspiel D&B vs T&L&B,  Becker 2015   Gelöst in 0.02s/7; Gelöst: 15
16. ERET 016 - Pos. Opfer,  Bischoff-Fressinet, Dt. Bundesliga 2008   Gelöst in 0s/7; Gelöst: 16
17. ERET 017 - Koenigsangriff,  Boudy-Szabo, Varna 1979   Gelöst in 6.68s/17; Gelöst: 17
18. ERET 018 - Entwicklung,  Bromberger-Areshchenko, Dt-BL 2010   Gelöst in 0.11s/10; Gelöst: 18
19. ERET 019 - Endspiel L&L&S&S vs L&S&B,  Broni 1970   Gelöst in 0s/7; Gelöst: 19
20. ERET 020 - Laeuferopfer h7,  Burmakin-Ziatdinov, Moskau 2009   Gelöst in 0.02s/7; Gelöst: 20
21. ERET 021 - Prophylaxe,  Carlsen-Nakamura, London 2010   Gelöst in 5.76s/17; Gelöst: 21
22. ERET 022 - Freibauer,  Carlsen-Nakamura, Medias 2011   Gelöst in 0.56s/16; Gelöst: 22
23. ERET 023 - Rochadeangriff,  Copie-Patrici, CorrGame 1986   Gelöst in 0s/7; Gelöst: 23
24. ERET 024 - Entwicklung,  Critter-Komodo, CompGame 2016   Gelöst in 0s/7; Gelöst: 24
25. ERET 025 - Endspiel T&L vs T&L,  Critter-Naum, CompGame 2015   Gelöst in 0s/7; Gelöst: 25
26. ERET 026 - Springeropfer,  Degraeve-Aseev, St.Petersburg 1996   Gelöst in 0.19s/12; Gelöst: 26
27. ERET 027 - Zugzwang,  DeLimburg 1900   Gelöst in 0.05s/7; Gelöst: 27
28. ERET 028 - Vergifteter Bauer,  Diep-DeepJunior, WCCC 2004   Gelöst in 0s/7; Gelöst: 28
29. ERET 029 - Qualitaetsopfer,  Dolinskij-Tosic, Moskau 1999   Gelöst in 0s/7; Gelöst: 29
30. ERET 030 - Initiative,  Dutra-Alfredo, CorrGame 2008   Gelöst in 0s/7; Gelöst: 30
31. ERET 031 - Endspiel L vs L,  Estrin-Ivashin, CorrGame 1947   Gelöst in 0s/7; Gelöst: 31
32. ERET 032 - Zugzwang,  Fahrni 1922   Gelöst in 0s/7; Gelöst: 32
33. ERET 033 - Initiative,  Fedorchuk-Sokolov, Khanty-Mansiysk 2009   Gelöst in 0s/7; Gelöst: 33
34. ERET 034 - Laeuferpaar,  Fischer-Incutto, Mar Del Plata 1960   Gelöst in 0s/7; Gelöst: 34
35. ERET 035 - Qualitaetsopfer,  FischerW-Siigur, CorrGame 2009   Gelöst in 0s/7; Gelöst: 35
36. ERET 036 - Endspiel T&L&S vs T&L&S,  Fritz 1965   Gelöst in 0s/7; Gelöst: 36
37. ERET 037 - Zugzwang,  Garcia 2000   Gelöst in 0s/7; Gelöst: 37
38. ERET 038 - Endspiel S&S vs L&B,  Garcia 2000   Gelöst in 0s/7; Gelöst: 38
39. ERET 039 - Entwicklung,  Gelfand-Ivanchuk, Dagomys 2009   Gelöst in 0.05s/7; Gelöst: 39
40. ERET 040 - Koenigssicherheit,  Gerasin-Dronov, CorrGame 1991   Gelöst in 2.43s/17; Gelöst: 40
41. ERET 041 - Springeropfer,  Gerber-Georgiev, Luzern 2003   Gelöst in 0.05s/7; Gelöst: 41
42. ERET 042 - Endspiel S vs L,  Glaurung-Rybka (Var), CompGame 2009   Gelöst in 0s/7; Gelöst: 42
43. ERET 043 - Springeropfer,  Glek-Ulibin, Wien 1998   Gelöst in 0.05s/7; Gelöst: 43
44. ERET 044 - Linienoeffnen,  Gligoric-Ghitescu, Zagreb 1970   Gelöst in 0.05s/7; Gelöst: 44
45. ERET 045 - Initiative,  Gouw-BlancoG, CorrGame 2001   Gelöst in 0s/7; Gelöst: 45
46. ERET 046 - Endspiel T&S vs T&S,  Grandelius-Raznikov, Albena 2011   Gelöst in 0s/7; Gelöst: 46
47. ERET 047 - Entlastung,  Grischuk-Aronian, Ohrid 2009   Gelöst in 0s/7; Gelöst: 47
48. ERET 048 - Starke Felder,  Grivas-Svetushkin, Glifada 2001   Gelöst in 0s/7; Gelöst: 48
49. ERET 049 - Entwicklung,  Gufeld-Ivanovic, Socchi 1979   Gelöst in 9.33s/20; Gelöst: 49
50. ERET 050 - Koenigsangriff,  Hansen-Barua, Biel 1993   Gelöst in 0s/7; Gelöst: 50
51. ERET 051 - Verteidigung,  Harding-Zavanelli, CorrGame 1985   Gelöst in 0.30s/12; Gelöst: 51
52. ERET 052 - Starke Felder,  HarEven-Kofidis, CorrGame 1996   Gelöst in 0s/7; Gelöst: 52
53. ERET 053 - Pos. Opfer,  Ivanisevic-Nisipeanu, Istanbul 2003   Gelöst in 0.05s/7; Gelöst: 53
54. ERET 054 - Endspiel D&T&L vs D&T&L,  Jussupow-Kasparov, Linares 1990   Gelöst in 0s/7; Gelöst: 54
55. ERET 055 - Laeuferopfer h6,  Kamsky-Lautier, Dortmund 1993   Gelöst in 0.05s/7; Gelöst: 55
56. ERET 056 - Zwischenzug,  Karjakin-Kramnik, Moskau 2010   Gelöst in 0.05s/7; Gelöst: 56
57. ERET 057 - Abtausch,  Kashlyak-Vandermeulen, CorrGame 2008   Gelöst in 0.02s/7; Gelöst: 57
58. ERET 058 - Endspiel T&L vs T&L,  Kasparian 1954   Gelöst in 0s/7; Gelöst: 58
59. ERET 059 - Endspiel D&L vs D&T,  Komodo-Stockfish, CompGame 2014   Gelöst in 0.08s/11; Gelöst: 59
60. ERET 060 - Koenigsangriff,  Kortschnoj-Andersson, Ubeda 1997   Gelöst in 0s/7; Gelöst: 60
61. ERET 061 - Koenigsangriff,  Kortschnoj-Shirov, Madrid 1996   Gelöst in 0s/7; Gelöst: 61
62. ERET 062 - Starke Felder,  Kotronias-Davies, Gausdal 1094   Gelöst in 0.02s/7; Gelöst: 62
63. ERET 063 - Verteidigung,  Kovacs-Sperkin, CorrGame 2012   Gelöst in 0.36s/12; Gelöst: 63
64. ERET 064 - Endspiel T vs T,  Ktulu-Hiarcs (Var), CompGame 2005   Gelöst in 0s/7; Gelöst: 64
65. ERET 065 - Zwischenzug,  Kuligowski-Knaak, Slupsk 1982   Gelöst in 0.05s/7; Gelöst: 65
66. ERET 066 - Endspiel T&S vs L&S,  Marwitz 1965   Gelöst in 0s/7; Gelöst: 66
67. ERET 067 - Raeumung,  Moisieev-Simagin, Moskau 1951   Gelöst in 2.12s/18; Gelöst: 67
68. ERET 068 - Offene Linie,  Morozevich-Kogan, London 1994   Gelöst in 0.02s/7; Gelöst: 68
69. ERET 069 - Koenigsangriff,  Morozevich-Petursson, London 1994   Gelöst in 0s/7; Gelöst: 69
70. ERET 070 - Starke Felder,  Motylev -Li, China 2010   Gelöst in 0.03s/10; Gelöst: 70
71. ERET 071 - Ablenkung,  Myers-Baufays, CorrGame 2010   Gelöst in 0s/7; Gelöst: 71
72. ERET 072 - Zentralisierung,  N.N.-N.N. (Analyse)   Gelöst in 0.05s/7; Gelöst: 72
73. ERET 073 - Mobilitaet,  N.N.-N.N. (Analyse)   Gelöst in 0.05s/7; Gelöst: 73
74. ERET 074 - Endspiel T&S vs T&S,  N.N.-N.N. (Analyse)   Gelöst in 0s/7; Gelöst: 74
75. ERET 075 - Festung,  N.N.-N.N. (Analyse)   > 15s.
76. ERET 076 - Entwicklung,  Nakamura-VanWely, Amsterdam 2010   Gelöst in 0s/7; Gelöst: 75
77. ERET 077 - Rochadeangriff,  Negi-Volokitin, ESP 2008   Gelöst in 0s/7; Gelöst: 76
78. ERET 078 - Freibauer,  Niedermayr-Bachschmidt, CorrGame 1968   Gelöst in 0s/7; Gelöst: 77
79. ERET 079 - Damenopfer,  Nijboer-Winants, WijkAanZee 1988   Gelöst in 0.02s/7; Gelöst: 78
80. ERET 080 - Raeumung,  Nirvana-Hannibal, CompGame 2015   Gelöst in 0s/7; Gelöst: 79
81. ERET 081 - Koenigsangriff,  Nunn-Ward, GB 1998   Gelöst in 0s/7; Gelöst: 80
82. ERET 082 - Endspiel S vs L,  Nurmamed.-Volovici (Var), CorrGame 1960   Gelöst in 0s/7; Gelöst: 81
83. ERET 083 - Abtausch,  Orlov-Georgiev, Belgrad 1988   Gelöst in 0.05s/7; Gelöst: 82
84. ERET 084 - Koenigsangriff,  Persson-Teichmeister, CorrGame 2009   Gelöst in 0s/7; Gelöst: 83
85. ERET 085 - Abtausch,  Petrosian-Bronstein, UdSSR-Ch 1967   Gelöst in 0s/7; Gelöst: 84
86. ERET 086 - Qualitaetsopfer,  Petrosian-Fischer, BuenosAires 1971   Gelöst in 0s/7; Gelöst: 85
87. ERET 087 - Raeumung,  Petrosian-Larsen, PalmaDeMallorca 1968   Gelöst in 0s/7; Gelöst: 86
88. ERET 088 - Endspiel S vs S,  Razuvaev-Ostojic,  1988   Gelöst in 0s/7; Gelöst: 87
89. ERET 089 - Unterverwandlung,  Rumänien 1977   Gelöst in 0.13s/14; Gelöst: 88
90. ERET 090 - Endspiel L vs L,  Salai 2011   Gelöst in 0s/7; Gelöst: 89
91. ERET 091 - Endspiel B vs B,  Shredder-Fritz (Var), CompGame 2007   Gelöst in 0s/9; Gelöst: 90
92. ERET 092 - Vergifteter Bauer,  Shredder-Ruffian, CompGame 2005   Gelöst in 0.02s/7; Gelöst: 91
93. ERET 093 - Endspiel T&S&B vs T&T&L,  Simkhovich 1923   > 15s.
94. ERET 094 - Damenopfer,  Sirobaba-Chaika, CorrGame 2011   Gelöst in 0s/7; Gelöst: 92
95. ERET 095 - Damenopfer,  Smagin-Sahovic, Biel 1990   Gelöst in 0s/7; Gelöst: 93
96. ERET 096 - Endspiel D vs T&L&S,  Sochnev 2009   Gelöst in 0.05s/7; Gelöst: 94
97. ERET 097 - Endspiel L vs L,  Spassky-Barne, SanJuan 1974   Gelöst in 0s/7; Gelöst: 95
98. ERET 098 - Endspiel T vs T,  Stockfish-Komodo, CompGame 2015   Gelöst in 0s/7; Gelöst: 96
99. ERET 099 - Endspiel D&T vs D&T&S,  Stuovsky-Movsesian, EU-Ch 2001   Gelöst in 0.05s/7; Gelöst: 97
100. ERET 100 - Initiative,  Sutovsky-Agrest, Harplinge 1998   Gelöst in 0s/7; Gelöst: 98
101. ERET 101 - Entwicklung,  Tronhjem-Christensen, CorrGame 1984   Gelöst in 0s/7; Gelöst: 99
102. ERET 102 - Koenigsangriff,  Tumurhuyag-Uhlmann, Novi Sad 1990   Gelöst in 0s/7; Gelöst: 100
103. ERET 103 - Zugzwang,  VanEssen 2004   Gelöst in 2s/23; Gelöst: 101
104. ERET 104 - Entwicklung,  VanOosterom-Timmerman, CorrGame 1982   Gelöst in 1.11s/15; Gelöst: 102
105. ERET 105 - Starke Felder,  Vlasveld-Neagu, CorrGame 2014   Gelöst in 0.03s/9; Gelöst: 103
106. ERET 106 - Koenigssicherheit,  Wahls-Chandler (Var), NoviSad 1990   Gelöst in 0.94s/16; Gelöst: 104
107. ERET 107 - Verteidigung,  Wiacek-Kolarik, CorrGame 1999   Gelöst in 0.02s/7; Gelöst: 105
108. ERET 108 - Endspiel T&L vs T&L,  Winckelmann-Bastos, CorrGame 2003   Gelöst in 0s/7; Gelöst: 106
109. ERET 109 - Entlastung,  Xie-Bischoff, Baden-Baden 1992   Gelöst in 1.95s/15; Gelöst: 107
110. ERET 110 - Freibauer,  Ye-Lutz, Batumi 2001   Gelöst in 0s/7; Gelöst: 108
111. ERET 111 - Festung,  Zueger-Kindermann(Var), Muenchen 1989   Gelöst in 2.54s/21; Gelöst: 109

Ergebnis: 109 aus 111 = 98.1%. Durchschnittszeit = 0.35s / 8.53

[corres]

The speed of a net is such an important parameter than the knowledge stored in NN. A "clever" but too slow net is not better then a fast but "stupid" net. The task of net-makers is for making nets with optimal speed and knowledge.

I would say that it is rather YOUR Task, not NN maker's, to make sure you have the fastest possible GPU to fully UTILIZE the large NN.
Its rather silly to use a slow Graphic Card for Testing and then say that large NN is bad !
You neglect the importance of Hardware for testing NNs.

In your opinion if a net is finished what a user can be with it?
If you are curious to know how the net behave on your powerful system you must make tests with it, do not you think so?
I do not make a more expensive stronger system making tests for your kindness.
In every cases I write down my system and the used parameters so everybody who think himself being an expert can value my results and that one can imagine how behave the tested net on his own system.
I think only the laziness can hinder others to retest the net on his stronger system.

[corres]

...
Yes I don't know why the testing sites only play super lightning blitz tournament on only default settings and not the best settings except for me to rate NN engines. I am glad I do not. I guess they stupidly assume that all NN are created equal in regards to the default settings.

If you have a lot of time to waste please, make a line of tests to optimize the parameters and the another line of tests to show the relations among nets.

I have found the bigger nets to be better at shorter time controls and longer time controls. NPS is a very poor gauge on what is the best NN at any time control. I have show this in my testing.
...

Where are the arguments?
The speed of a net is such an important parameter than the knowledge stored in NN. A "clever" but too slow net is not better then a fast but "stupid" net. The task of net-makers is for making nets with optimal speed and knowledge.

Lets see where Sergio 40x512 stands in regards to the other engines and NN.

Yes testing with correct settings. That is to stupid. I guess we know you are the problem. To bad I am here to show you are full of shit.
Testing with correct settings is self evident.

Who hinder you for making such a "correct test" what you think being correct? I am not, it is sure.
So come on boy, let's see the results!

[corres]

...
I have very very very good news .
I've tested the frist 40x512 net with the ERET test suite.
First result with 8GB RAM and a limit of 15 seconds per move: 59 of 111 points.
Second result with 64 GB RAM and a limit of 300 seconds per move: 82 of 111 points.
This is an improvement of 23 points holy shit
83 / 111 = 74.7%.
Mean time = 36.40s / 9.97
Usually when doing the same test as described before with the smaller nets, the improvement was only ~2 points more than before.
This is the first net which have the capability to solve the complete ERET Test Suite after it is trained.

The explanation is very simple:
For exhausting a smaller net it needs more fewer time than a big net.
15 sec may be enough for that smaller net to exhaust it but for the 512x40 even the 300 sec may be too few.

[mwyoung]

A 512x40 test is running here:

https://www.comp.nus.edu.sg/~sergio-v/test/

Thank you very much. It looks much better running LC0 Sergio 40x512 net with my 2x RTX 2080 Ti.

This is amazing

Hello Hai. What speed do you get in this FEN position. I'm asking because i want to compare the speed difference. I'm running 2x RTX 2080 Super OC (ASUS ROG STRIX) Identical cards.

[d]1q6/1r2k1p1/4pp1p/1P1b1P2/3Q4/7P/4B1P1/2R3K1 w

Analyse for 1 minute with 512x40 net.

I'm using following settings tweaked from the defaults:

<------->
Roundrobin
(backend=cudnn-fp16,gpu=0),(backend=cudnn-fp16,gpu=1)
CPuct: 3.4
CPuctBase: 10000
MiniBatchSize: 512
<------->

Tablebases disabled.

Assuming you have two identical cards try same settings or multiplexing if they're not intentical and run that test for speed comparison

1 minute search:

1q6/1r2k1p1/4pp1p/1P1b1P2/3Q4/7P/4B1P1/2R3K1 w - - 0 1

Analysis by Lc0 v0.23.2+git.c8d9095:

1.Qg4 exf5 2.Qxg7+ Ke6 3.Qg6 Qe5 4.Qe8+ Re7 5.Qg8+ Rf7 6.Re1 Kd6 7.b6 Rd7 8.Qf8+ Re7 9.Qb8+ Ke6 10.Qc8+ Kd6 11.Kh1 Bxg2+ 12.Kxg2 Qe4+ 13.Kf2 Qe3+ 14.Kf1
White is better: +/- (1.14) Depth: 12/35 00:01:06 329kN
(12.01.2020)

2x 2080 Super output: speed 5544 N/s after about one minute.

Hi Nordlandia,

we have nearly the same N/s and kN.

I'm using default settings.
And also:
Roundrobin
(backend=cudnn-fp16,gpu=0),(backend=cudnn-fp16,gpu=1)
+ 64 GB RAM

You have 2x 2080 Super and OC.
I have 2x 2080 Ti without Super and without OC.
Also I use power limit of 95% per GPU.
Also I use temperature limit of 70 per GPU.
So I've got lower power consumption, lower heat and it's also really quiet .
But GPU load is jumping between 60 to 90% on both GPUs when running your position.

When I run chess start position I have GPU load of 93% on both GPUs.
And CPU heat is 50.

I guess the question is does this network 512x40 need 1, 2,3,4....High end graphic cards. Assuming it is a good net to be the best. Have you run any game test with this Fat Network. I will download it for a full round of testing. Have we now reached the limits of 1 RTX graphics card?

I can test if this Fat Network can top the best on 1 RTX....

I'm interested in chess analysis and not in engine-super-lightning-bullet-tc-chess .

Like for every big net, at the beginning you should have 2x RTX GPUs, 2080 Ti if possible.
Like for every big net, later one RTX GPU, 2080 Ti if possible, will be enough.
-30x384-T60-2, 131 MB, : Ongoing, at 3rd LR (0.002) - That explains why 1 RTX GPU is enough.
-40x512, 318 MB, already started to train = less knowledge at the moment = to compensate it you sould use 2x RTX GPUs. But it also benefit much more from 3 or 4 RTX GPUs than the other nets.

-I can do now much longer analysis, when using the LC0 Sergio 40x512 net.
-I need much lesser RAM when running the 40x512 net.
-It is also possible to run it much longer than a 30x384 net before I run out of memory with my 64 GB RAM.

I expect the 40x512 net to get ~60 points in the ERET Test, because it's an untrained net at the moment.
Of course when using longer time control, the result will be better.
I have tested it also for analysis and games playing style.
It can already find some very difficult moves when using it on some fen positions to analyse. That's really amazing.
Strategically it is weaker at the moment and the positional style is also weaker at the moment.
But tactics are better .

Yes I don't know why the testing sites only play super lightning blitz tournament on only default settings and not the best settings except for me to rate NN engines. I am glad I do not. I guess they stupidly assume that all NN are created equal in regards to the default settings.
I have found the bigger nets to be better at shorter time controls and longer time controls. NPS is a very poor gauge on what is the best NN at any time control. I have show this in my testing.

And I assume you know that every position in a chess game is a series of positions to analyze. And the engine that is most correct general wins the game. That is why we test. General the strongest chess playing engine will give you the best analysis.

Lets see where Sergio 40x512 stands in regards to the other engines and NN.

I have very very very good news .
I've tested the frist 40x512 net with the ERET test suite.
First result with 8GB RAM and a limit of 15 seconds per move: 59 of 111 points.
Second result with 64 GB RAM and a limit of 300 seconds per move: 82 of 111 points.

This is an improvement of 23 points holy shit
83 / 111 = 74.7%.
Mean time = 36.40s / 9.97
Usually when doing the same test as described before with the smaller nets, the improvement was only ~2 points more than before.
This is the first net which have the capability to solve the complete ERET Test Suite after it is trained.

My first round of testing is giving the Sergio 512 network a jump of 17 elo over the sergio 1705 network in early testing.