Lc0 v0.28.2 on GTX 1050ti - Looking for best Settings

[Eduard]

Looking for best Settings for Lc0 v0.28.2 and GTX 1050Ti? A lot has happened with Lc0, so that I no longer know the best settings for my graphics card + v0.28.2. Which settings for NNCasheSize, Minibatchsize, MaxPrefetch, Cpuct, etc... are recommended?

Testposition with Default-Settings on GTX 1050Ti (3600 kns with 15x192 Network):

[fen]5b2/p1p4k/2Pq2pp/4N3/2Bp1r2/PR6/K3Q3/8 w - - 0 1[/fen]


Analysis by Lc0 v0.28.2 with network 771721 (15x192):

46.Bd3 Rf6 47.Nxg6 Rxg6 48.Qf2 Bg7 49.Qf5 Qf6 50.Rb2 Qxf5 51.Bxf5 Bf6 52.Rb7 Be5 53.Rxa7 Kg7 54.Bxg6 Kxg6 55.a4 h5 56.Rb7 d3 57.Rb1 h4 58.a5 h3 59.a6 Bd4 60.Rh1 d2 61.Kb3 h2
White is clearly better: +/- (1.49) Depth: 15/32 00:00:25 91kN, tb=29

Looks good here.

[Eduard]

I remember. Here are some positions that were difficult for Lc0 with my GTX 1050Ti 1,5 years ago. Now tested with network 611245.pb.gz (30x380) with the following settings (lc0.config). The 15x190 network is weaker than the 30x384 network on my hardware too.

--threads=2
--minibatch-size=128
--max-prefetch=16
--nncache=2000000
--cpuct-base=10500.000000
--cpuct-base-at-root=10500.000000
--policy-softmax-temp=1.600000
--fpu-value=0.420000
--move-overhead=30000
--time-manager=legacy(slowmover=0.458000,midpoint-move=24.00,steepness=4.00)
--syzygy-paths=G:\3456men

Diagram 1:
[fen]1k6/1P3p2/5P2/pq6/1pp1BP1P/6P1/1P1r4/R3R2K w - - 0 1[/fen]


Analysis by Lc0 v0.28.2 30x84:

31.Bg2 Rxg2 32.Kxg2 b3 33.Rad1 Kxb7 34.Re7+ Kc8 35.Rxf7 c3 36.Rf8+ Kc7 37.Rf7+ Kb6 38.Rd6+ Kc5 39.Re6 Qb8 40.Re5+ Kc4 41.Re4+ Kd3 42.Rfe7 cxb2 43.Re3+ Kc4 44.R3e4+ Kb5 45.R7e5+ Qxe5
The position is equal: = (-0.11) Depth: 13/30 00:00:37 24kN
(---)
31.Bg2 Rxg2 32.Kxg2 a4 33.Rad1 Qxb7+ 34.Kh2 Qf3 35.Kh3 Qc6 36.Re7 Kc8 37.Kh2 Qxf6 38.Rdd7 Qc6 39.Ra7 a3 40.bxa3 b3 41.Rxf7 Kb8 42.Rae7 Qc8 43.Re5 c3 44.Rb5+ Ka8 45.Ra5+ Kb8 46.Rb5+
The position is equal: = (-0.10) Depth: 17/47 00:03:31 174kN, tb=2

In the past Lc0 played here on my GPU Bf3? and losed.

Diagram 2:
[fen]1r2r1k1/5pb1/4q1p1/3Qp1p1/1p2P3/pP4PP/P1R2PK1/B1R5 w - - 0 1[/fen]


Analysis by Lc0 v0.28.2 30x84:

44.Rc6 Qxd5 45.exd5 Red8 46.d6 f6 47.Rd1 Rd7 48.f4 gxf4 49.gxf4 Kf7 50.fxe5 fxe5 51.Rc5 Re8 52.Rb5 Ke6 53.Bxe5
The position is equal: = (-0.10) Depth: 8/19 00:00:08 4897N
44.Rc6 Qxd5 45.exd5 Red8 46.d6 f6 47.Rd1 Rd7 48.f4 gxf4 49.gxf4 Kf7 50.fxe5 fxe5 51.Rc5 Re8 52.Rb5 Ke6 53.Bxe5 Bxe5
(---)
44.Rc6 Qxd5 45.exd5 Red8 46.d6 f6 47.Rd1 Rd7 48.f4 gxf4 49.gxf4 Kf7 50.fxe5 fxe5 51.Rc5 Re8 52.Rb5 Re6 53.Rxb4 Rexd6 54.Rxd6 Rxd6 55.Rb7+ Kf6 56.Bxe5+ Kxe5 57.Rxg7 Rd2+ 58.Kf3 Rxa2 59.Rxg6
The position is equal: = (-0.09) Depth: 12/33 00:01:30 71kN

In the past, Lc0 played Qxe6?? and losed.

Diagram 3:
[fen]8/p5p1/1p4k1/1P1PQp2/Pq4p1/6P1/4K2P/8 b - - 0 1[/fen]


Analysis by Lc0 v0.28.2 30x84:

66...Qxa4 67.Qe6+ Kg5 68.Qe7+ Kg6 69.Qe6+
The position is equal: = (-0.03) Depth: 11/33 00:00:22 33kN, tb=179
66...Qe4+ 67.Qxe4 fxe4 68.d6 Kf6 69.d7 Ke7 70.d8Q+ Kxd8 71.Ke3 Kd7 72.Kxe4 Kd6 73.Kd4 g6 74.Kc4 Ke5 75.Kd3 Kd5 76.Ke3 Kc5 77.Kf4 Kb4 78.Kxg4 Kxa4 79.h4 Kxb5 80.Kf4 Kc5 81.Ke5 a5
Black is better: -/+ (-0.86) Depth: 11/33 00:00:23 37kN, tb=318
(---)
66...Qe4+ 67.Qxe4 fxe4 68.d6 Kf6 69.d7 Ke7 70.d8Q+ Kxd8 71.Ke3 Kd7 72.Kxe4 Kd6 73.Kd4 g6 74.Kc4 Ke5 75.Kd3 Kd5 76.Ke3 Kc5 77.Kf4 Kb4 78.Kxg4 Kxa4 79.h4 Kxb5 80.Kg5 a5 81.g4 a4 82.Kf4 a3
Black is better: -/+ (-0.99) Depth: 13/34 00:00:24 42kN, tb=554

In the past, Lc0 needed a few minutes to see 66...Qe4+ here and losed with white pieces. Now Lc0 needs only 23s on my hardware. Good progress!

Diagram 4:
[fen]3qkr2/pp1b1pnr/4p1p1/3pP3/5QBP/2P5/PP1N1P2/2KR3R b - - 0 1[/fen]


Analysis by Lc0 v0.28.2 30x84:

23...Rxh4 24.Rxh4 Qxh4 25.Ne4 Nf5 26.Nf6+ Kd8 27.Qb4 Qh6+ 28.Kb1 Bc6 29.c4 Kc8 30.cxd5 exd5 31.Rc1 Rh8 32.Qd6 Qf4 33.f3 Rd8 34.Rxc6+ bxc6
White is slightly better: +/= (0.43) Depth: 12/30 00:01:04 42kN
23...Nh5 24.Qh2 Qa5 25.Kb1 Qa4 26.Qg1 Nf4 27.b3 Qa6 28.Nf3 Rfh8 29.Ng5 Rxh4 30.Rxh4 Rxh4 31.Qg3 Ne2 32.Qxh4 Nxc3+ 33.Kc2 Nxd1 34.Qh8+ Ke7
White is slightly better: +/= (0.45) Depth: 12/30 00:01:07 45kN

In the past, Lc0 on my hardware here took more than 2 minutes and played Rxh4?? Now only 67s to find Nh5. That's progress!

Diagram 5:
[fen]1q2r3/5ppk/2b2r1p/p3nN2/2P1p3/BP1nQ1NP/P4PP1/1R1R2K1 w - - 0 1[/fen]


Analysis by Lc0 v0.28.2 30x84:

30.Be7 Re6 31.Bc5 Rg6 32.Bd4 Nf4 33.Ba7 Qc7 34.Bb6 Qb8 35.Ba7
The position is equal: = (0.05) Depth: 12/37 00:00:52 37kN
30.Nxg7 Rg8 31.Rxd3 Rxg7 32.Rd5 Bxd5 33.cxd5 Nf3+ 34.Kh1 Rxg3 35.fxg3 Qxg3 36.Qxe4+ Kh8 37.Bd6 Rxd6 38.Qxf3 Qxf3 39.gxf3 Rxd5 40.Rb2 Kg7 41.Kg2 Kf6 42.Kg3 Rg5+ 43.Kf4 Rh5 44.Rh2 a4
White is slightly better: +/= (0.31) Depth: 12/37 00:00:55 38kN

In less than 60s on my hardware is a very good value for Lc0 around Nxg7! to find.

Diagram 6:
[fen]1br2rk1/4q1p1/p3pn1p/Ppp1N3/1P2Pn2/4R1NP/1B1R1PP1/3Q2K1 b - - 0 1[/fen]


Analysis by Lc0 v0.28.2 30x84:

25...c4 26.Nd7 Nd3 27.Rdxd3 cxd3 28.Nxf8 Qxb4 29.Bxf6 d2 30.Bc3 Rxc3 31.Qxd2 Qb1+ 32.Re1 Rd3 33.Qe2 Qb4 34.Nxe6 Rd2 35.Qg4
White is slightly better: +/= (0.58) Depth: 10/28 00:00:17 7117N
25...Qa7 26.Qf3 Bxe5 27.Bxe5 Ng6 28.Bd6 Nd5 29.Bxf8 Rxf8 30.Qxf8+ Nxf8 31.exd5 exd5 32.Rxd5 cxb4 33.Rd8 Qc7 34.Ree8 b3 35.Rxf8+ Kh7 36.Rh8+ Kg6 37.Rc8 Qd7 38.Rhd8 Qe7 39.Rd1
The position is equal: = (0.21) Depth: 10/29 00:00:18 7626N
(---)
25...Qa7 26.Qf3 Bxe5 27.Bxe5 Ng6 28.Bd6 Nd5 29.Bxf8 Rxf8 30.Qxf8+ Nxf8 31.exd5 exd5 32.Rxd5 cxb4 33.Rd8 Qc7 34.Ree8 b3 35.Rxf8+ Kh7 36.Rh8+ Kg6 37.Rc8 Qxc8 38.Rxc8 b2 39.Rc6+
The position is equal: = (0.08) Depth: 11/31 00:00:21 9045N

25...c4? losing but Qb7 or Qa7 saves the game. Just 21s of Lc0 with net 611245.pb.gz is awesome! Lc0 has never been so good on my hardware.

I am positively surprised because graphics cards are very expensive, and for me Lc0 only makes sense if it costs less or is tested with cheaper graphics cards.