11258-32x4-se distilled network released

[dkappe]

11258-32x4-se distilled network released. Links and details on the wiki page: https://github.com/dkappe/leela-chess-w ... d-Networks

120x10 cooking now. Maybe 144x11 after.

[Eduard]

Thank you! I will test of course. It is fun. Such fast networks are optimal for mobile devices. Is there a chance that Lc0 will eventually run on Android?

[dkappe]

Thank you! I will test of course. It is fun. Such fast networks are optimal for mobile devices. Is there a chance that Lc0 will eventually run on Android?

I know someone was tinkering with an iOS version, not sure about android.

[Eduard]

Here are some analyzes of networks 104x9 and 32x4. You can see the difference. All in all, I would rather take 32x4. It is faster,
and there are not always only best moves to play.

Hardware = i3 2x2,4 GHz x64
GUI = Fritz 15

In following position, best move is 31. Rad1!

Kramnik,V - Howell,D 1-0, 43rd Olympiad Batumi 2018 Open
[d]5k2/pp2pN2/2n2rrp/2p4Q/7p/1Bq5/P4PP1/R3R1K1 w - - 0 1

Analysis by Lc0 v0.20.2-rc1 104x9:

31.Tac1 Dd2 32.Dxh4 Sd4 33.Se5 Txg2+ 34.Kxg2 Sxb3
+/- (1.23) Tiefe: 5/15 00:00:11 0kN
31.Tad1 c4 32.Te3 Db2 33.Te2 Dc3 34.Lc2 Txg2+ 35.Kxg2 Txf7
+- (1.67) Tiefe: 6/15 00:00:13 0kN
31.Tad1 h3 32.Ld5 Dc2 33.Dxh3 Dxf2+ 34.Kh2 Tf4 35.Se5 Th4 36.Sxg6+ Kg7
+/- (1.34) Tiefe: 6/15 00:00:18 1kN
31.Tad1 h3 32.Ld5 Dc2 33.Dxh3 Dxf2+ 34.Kh1 Tf4 35.Dc8+ Kg7 36.Dh8#
+/- (1.33) Tiefe: 6/16 00:00:19 1kN

Analysis by Lc0 v0.20.2-rc1 32x4:

31.Se5 Sxe5 32.Dxe5 Txg2+ 33.Kxg2 Df3+ 34.Kg1 Dxf2+ 35.Kh1 Df3+ 36.Kh2 Df2+ 37.Kh3 Df3+ 38.Kh2 Df2+ 39.Kh3 Df3+ 40.Kxh4 Tf4+ 41.Dxf4+ Dxf4+ 42.Kh3 Df3+ 43.Kh2 Df4+ 44.Kh1 Dh4+ 45.Kg2 Dg4+ 46.Kf2 c4 47.Lc2 Df4+ 48.Ke2 De5+ 49.Kd2
+/- (0.86) Tiefe: 19/39 00:02:04 160kN
31.Se5 Sxe5 32.Dxe5 Txg2+ 33.Kxg2 Df3+ 34.Kg1 Dxf2+ 35.Kh1 Df3+ 36.Kh2 Df2+ 37.Kh3 Df3+ 38.Kh2 Df2+ 39.Kh3 Df3+ 40.Kxh4 Tf4+ 41.Dxf4+ Dxf4+ 42.Kh3 Df3+ 43.Kh2 Df4+ 44.Kh1 Dh4+ 45.Kg2 Dg4+ 46.Kf2 c4 47.Lc2 Df4+ 48.Ke2 De5+ 49.Kd2
+/- (0.86) Tiefe: 19/39 00:02:09 165kN
31.Tad1 h3 32.Ld5 Dc2 33.Dxh3 Dxf2+ 34.Kh2 Df4+ 35.Kg1 Df2+ 36.Kh2 Df4+ 37.Kh1 Txf7 38.Lxf7 Dxf7 39.Tf1 Tf6 40.Kg1 Kg7 41.Dg3+ Dg6 42.Dxg6+ Txg6 43.Td7 b5 44.Tc7 Td6
+- (1.72) Tiefe: 19/39 00:02:09 167kN
31.Tad1 h3 32.Ld5 Dc2 33.Dxh3 Dxf2+ 34.Kh2 Df4+ 35.Kg1 Df2+ 36.Kh2 Df4+ 37.Kh1 Txf7 38.Lxf7 Dxf7 39.Tf1 Tf6 40.Kg1 Kg7 41.Txf6 Dxf6 42.Dc8 Sd4 43.Dxc5 Se2+ 44.Kh2 Dh4#
+- (1.66) Tiefe: 19/39 00:02:14 175kN

In following position, white have a advantage after 1. Nh6+.

Nemeth,E - Fritz 80026 1-0, Blitz:10' Privat 2004
[d]rn1qrnk1/p4pp1/1p1pp3/6P1/2Pp1PN1/2PQ4/P5P1/2KR3R w - - 0 1

Analysis by Lc0 v0.20.2-rc1 32x4:

18.Dh3 Sg6 19.Txd4 Kf8 20.Dh7 Sc6 21.Td2 Sa5 22.f5 exf5 23.Sh6
=/+ (-0.61) Tiefe: 6/14 00:00:08 4kN
18.Sh6+ gxh6 19.Txh6 Sbd7 20.Tdh1 Sc5 21.Th8+ Kg7 22.Dxd4+ f6 23.T1h7+ Sxh7 24.gxf6+
= (-0.19) Tiefe: 6/14 00:00:08 4kN

Analysis by Lc0 v0.20.2-rc1 104x9:

18.Th2 Sbd7 19.Tdh1 f5 20.gxf6 Sxf6 21.Th8+ Kf7 22.Sxf6 Dxf6
= (-0.05) Tiefe: 5/10 00:00:09 1kN
18.Th2 Sbd7 19.Tdh1 f5 20.gxf6 Sxf6 21.Th8+ Kf7 22.Sxf6 Dxf6
= (-0.16) Tiefe: 5/11 00:00:10 1kN
18.Sh6+ gxh6 19.Txh6 Sbd7 20.Tdh1 f5 21.g4 Se5 22.fxe5 Dxg5+
= (-0.08) Tiefe: 5/11 00:00:10 1kN

In following position, white is better, yes. But 18...Qh4 or Qf6 loses instantly!

Rybka 4.1 x64 - Lc0 v0.20.1distilled 1-0, ASUS-PC, Blitz 7m+3s
[d]r2q1rk1/1p3pp1/p2b4/3p3b/3N2P1/3B3P/PPPQ1P2/R4RK1 b - - 0 1

Analysis by Lc0 v0.20.2-rc1 32x4:

18...Df6 19.Sf5 Lg6 20.Sxd6 Dxd6 21.Lxg6 Dxg6 22.Dxd5 Dxc2 23.Dxb7 Tab8 24.Dxa6 Txb2 25.Da3 Tfb8 26.Tac1 Dd2 27.Tcd1 Df4 28.De3 Dxe3
+/- (1.51) Tiefe: 13/34 00:03:43 253kN
18...Lxg4 19.hxg4 Dh4 20.f3 g6 21.Df2 Lg3 22.Dg2 Lf4 23.Se2 Le3+ 24.Tf2 Kg7 25.Dg3 Th8 26.Kg2 Lxf2 27.Kxf2 Df6 28.Tb1 Tae8 29.b3 Dg5 30.f4 Df6 31.g5 Db6+ 32.Kg2 Te3 33.Df2
+/- (1.42) Tiefe: 13/34 00:03:45 254kN

Analysis by Lc0 v0.20.2-rc1 104x9:

18...Df6 19.Sf5 Lg6 20.Sxd6 Dxd6 21.Lxg6 Dxg6 22.Tad1 Tad8 23.Dd3 Df6 24.c3 d4 25.cxd4 Td5 26.Tfe1 Tfd8 27.Te4
+/- (1.53) Tiefe: 12/27 00:06:23 68kN
18...Lxg4 19.hxg4 Dh4 20.f3 g6 21.Tf2 Tae8 22.Tg2 Kg7 23.Dc3 Kg8 24.Se2 Te3 25.Tf1 Tfe8 26.Tff2 Dh3 27.f4 Txe2 28.Txe2
+/- (1.16) Tiefe: 12/27 00:06:23 69kN

In following position, 55...Kxf4 loses instantly.

Lc0 v0.20.1 - Rybka 4.1 x64 0-1, ASUS-PC, Schnellschach 20m+5s 2019
[d]3k4/1R6/7p/p4P2/4Pr2/1P2K2p/8/8 w - - 0 1

Analysis by Lc0 v0.20.2-rc1 32x4:

55.Kxf4 h2 56.f6 Ke8 57.Tb8+ Kf7 58.Tb7+ Kxf6 59.Tb6+ Kg7 60.Tb7+ Kf8 61.Tb8+ Ke7 62.Tb7+ Kd8 63.Tb8+ Kc7 64.Th8 h1D 65.Th7+ Kb6 66.Td7 Dh3 67.Td6+
-+ (-1.64) Tiefe: 17/35 00:04:57 708kN, tb=9268
55.Tb8+ Kc7 56.Tb5 Th4 57.Tc5+ Kd6 58.Tc1 h2 59.Th1 Ke5 60.Kf3 Tf4+ 61.Kg3 Kxe4 62.Txh2 Txf5 63.Txh6 Kd3 64.Tb6 Kc3 65.Kg4 Td5 66.Kf3 Te5 67.Kf4 Te1 68.Tb5 Ta1 69.Ke3 a4
-/+ (-1.35) Tiefe: 17/35 00:04:58 713kN, tb=9541
55.Tb8+ Kc7 56.Tb5 Th4 57.Tc5+ Kd6 58.Tc1 h2 59.Th1 Ke5 60.Kf3 Tf4+ 61.Kg3 Kxe4 62.Txh2 Txf5 63.Txh6 Kd3 64.Tb6 Kc3 65.Kg4 Td5 66.Kf3 Te5 67.Kf4 Te1 68.Tb5 Ta1 69.Ke3 a4
-/+ (-1.36) Tiefe: 17/36 00:05:02 721kN, tb=9810

Analysis by Lc0 v0.20.2-rc1 104x9:

55.Kxf4 h2 56.f6 Ke8 57.Tb8+ Kf7 58.Tb7+ Kxf6 59.Tb6+ Kg7 60.Tb7+ Kf8 61.Tb8+ Ke7 62.Tb7+ Kd8 63.Tb8+ Kc7 64.Th8 h1D 65.Th7+ Kd6 66.e5+ Ke6 67.Ta7 Dh4+
-/+ (-1.28) Tiefe: 14/28 00:04:26 106kN, tb=179
55.Tb5 Th4 56.Td5+ Ke7 57.Td1 h2 58.Th1 Kf6 59.Kd4 h5 60.Kd5 Th3 61.Kd4 h4 62.Kd5 Txb3 63.Txh2 h3 64.Ta2 Tb5+ 65.Kd4 Tb4+ 66.Kd5 Tb5+ 67.Kd4
-/+ (-1.19) Tiefe: 14/28 00:04:27 108kN, tb=185
55.Tb5 Th4 56.Td5+ Ke7 57.Td1 h2 58.Th1 Kf6 59.Kd4 h5 60.Kd5 Th3 61.Kd6 h4 62.Kd5 Txb3 63.Txh2 h3 64.Ta2 Tb5+ 65.Kd4 Tb4+ 66.Kd5 Tb5+ 67.Kd4
-/+ (-1.21) Tiefe: 14/28 00:04:32 111kN, tb=209

[Eduard]

And here are few analyzes of new distilled network 120x10 (on i3 2x2,4 GHz):

[d]5k2/pp2pN2/2n2rrp/2p4Q/7p/1Bq5/P4PP1/R3R1K1 w - - 0 1
Analysis by Lc0 v0.20.2-rc1 120x10:

1.Ld5 Dd4 2.Tad1 Dxf2+ 3.Kh1 Dg3 4.Tf1 Kg7 5.Txf6 Txf6 6.Sd8
+/- (1.19) Tiefe: 6/13 00:00:19 1kN
1.Tad1 c4 2.Te3 Db2 3.Te2 Dc3 4.Lc2 Kxf7 5.Lxg6+ Txg6 6.Te4 Df6
+- (2.08) Tiefe: 6/13 00:00:20 1kN
1.Tad1 h3 2.Te3 Db2 3.Te2 Dc3 4.g3 Df3 5.Dxf3
+- (2.03) Tiefe: 6/13 00:00:25 1kN
1.Tad1 h3 2.Ld5 Dc2 3.Dxh3 Dxf2+ 4.Kh1 Kg7 5.Se5 Tg3 6.Dh2 Tg5
+- (1.71) Tiefe: 6/13 00:00:30 1kN
1.Tad1 h3 2.Ld5 Kg7 3.g3 Dc2 4.Tf1 Sd4 5.Se5
+/- (1.52) Tiefe: 6/13 00:00:35 1kN
1.Tad1 h3 2.Ld5 Kg7 3.Sd8 Sxd8 4.Txe7+ Sf7 5.Txf7+ Txf7 6.Lxf7 hxg2
+- (1.67) Tiefe: 7/14 00:00:38 1kN
1.Tad1 h3 2.Ld5 Kg7 3.Sd8 Sxd8 4.Txe7+ Sf7 5.Lxf7 Txg2+ 6.Kh1 Tfxf2 7.Lg8+
+- (1.73) Tiefe: 7/15 00:00:39 2kN



[d]r2q1rk1/1p3pp1/p2b4/3p3b/3N2P1/3B3P/PPPQ1P2/R4RK1 b - - 0 18
Analysis by Lc0 v0.20.2-rc1 120x10:

18...Df6 19.Sf5 Lg6 20.Sxd6 Dxd6 21.Lxg6 fxg6 22.Tad1 Tad8 23.c3 Tf3 24.Kg2 Tf4 25.f3 b5 26.a3 a5
+- (1.73) Tiefe: 9/22 00:02:25 17kN
18...Lxg4 19.hxg4 Dh4 20.f3 g6 21.Tf2 Tae8 22.Tg2 Df6 23.Sf5 gxf5 24.gxf5+ Kh7 25.Tg4 Tg8 26.Kg2 Txg4+
+/- (1.22) Tiefe: 9/22 00:02:29 18kN



[d]3k4/1R6/7p/p4P2/4Pr2/1P2K2p/8/8 w - - 0 1
Analysis by Lc0 v0.20.2-rc1 120x10:

1.Kxf4 h2 2.f6 Ke8 3.Tb8+ Kf7 4.Tb7+ Kxf6 5.Tb6+ Kg7 6.Tb7+ Kf8 7.Tb8+ Ke7 8.Tb7+ Kd8 9.Tb8+ Kc7 10.Th8 h1D 11.Th7+ Kd6 12.e5+ Ke6 13.Th8 Dh4+ 14.Kf3
-/+ (-1.37) Tiefe: 13/28 00:04:32 69kN, tb=127
1.Tb8+ Kc7 2.Tb5 h2 3.Tc5+ Kd6 4.Td5+ Ke7 5.Td1 Th4 6.Th1 Kf6 7.Kf3 h5 8.Kg3 Txe4 9.Txh2 Tg4+ 10.Kf3 Tg5 11.Ta2 Txf5+ 12.Kg3 Kg5 13.Ta4 Tb5 14.Ta3
-/+ (-1.36) Tiefe: 13/28 00:04:32 69kN, tb=127

[Guenther]

11258-32x4-se distilled network released. Links and details on the wiki page: https://github.com/dkappe/leela-chess-w ... d-Networks

120x10 cooking now. Maybe 144x11 after.

Thanks for those. I finally had some time to try one of them on my relatively new, but quite weak gpu!
(Actually I bought just before Leela appeared a new nvidia card with the goal just to be very silent and efficient - I am not into gaming)

Thus I installed a cheap GT 710, which has cuda support, but is ofc very slow for LC0.

With the 122*9se network nps increases may be by a factor of 5-6.
But this is not the only positive aspect. Due to passive cooling it always went too hot before even with an open case,
while it now runs around 10-15 degrees lower in non bullet games.

Currently testing against Laser 1.7.

[carldaman]

So, are these distilled networks also good for slower GPU cards? Which one (dn) is the strongest?
Thanks!

[dkappe]

So, are these distilled networks also good for slower GPU cards? Which one (dn) is the strongest?
Thanks!

It depends. There’s almost a step function at low nodes. 112x9-se is pretty good at 2+2. At 6+6 the larger distilled networks ot t35 may be better (i’ll Be doing a 144x11 and a 160x12 at some point). For a raspberry pi 3, a 48x5 might be best.

It’s all smarts vs speed. My advice it to pull down a few nets and run them vs a benchmark engine. Then pick the one that works for you.

For what I want — training games — a full size net at 100-200 nodes is plenty.

[Eduard]

It is good for a rough test 2m + 2s to play. But I do not see much sense in that, sorry.

Personally, I am more interested in a longer time control and for analysis mode.

When I'm on playchess.com I see more and more Lc0 playing there. But fewer and fewer Blitzgames with Lc0. Why? Well, most people are fascinated by the superior positional style of Lc0. Blitzgames are not good for that.

The private network DeusX, for example, is also being tested there. But never Blitzgames. I still like to kibitz when DeusX plays at 60m + 15s. I see the fascinating games of Lc0. Anyone who has ever seen the difference to blitzgames will understand me.

By the way, I like to play blitz too. With engines but also as a human being.

When I play bitz games with engines on playchess.com (I like it), then with big opening books and endgame databases. For that I take AB engines. Stockfish loves big books and syzygy.

But I will not play like that with Lc0.

[AdminX]

When I'm on playchess.com I see more and more Lc0 playing there. But fewer and fewer Blitzgames with Lc0. Why? Well, most people are fascinated by the superior positional style of Lc0. Blitzgames are not good for that.

The private network DeusX, for example, is also being tested there. But never Blitzgames. I still like to kibitz when DeusX plays at 60m + 15s. I see the fascinating games of Lc0. Anyone who has ever seen the difference to blitzgames will understand me.

I won't speak for everyone, but I play LTC Games because at this point in time Leela does not play to well at blitz. Still things are changing all the time so we will see.