Which engines are playing the opening well?

I am working on a new version of POLY 1.7 [ link ]. POLY is about creating and tuning Polyglot opening books. For the tuning part it is essential to use engines that understand the opening.

There is no doubt in my mind that the engine that plays the opening best is the Lc0. However Lc0 has a big disadvantage that it can only use one thread because most people have only one GPU installed. This contrary to the AB engines where the analysis can be divided over xx threads, 18 threads in my case with Stockfish 12 as the picture in the link shows.

Now SF12 is a great engine and the NNUE has improved its opening play, but still not good enough to my taste. And so I am looking for alternatives, engines that in your opinion play the opening well and can use multiple threads. Multi threads because analyzing 667K positions (the ProDeo book) is quite time consuming.

I would have to mention Komodo, but it's commercial. Shredder, too. I would consider Wasp and Winter among the free ones and maybe Ethereal, as well.

I was all forgotten about the OKE page

I have a NNUE version of Ethereal, maybe Pawel can recommend a personality of Rodent as it is doing extremely well between all the deep searchers.

Admin wrote:

I am working on a new version of POLY 1.7 [ link ]. POLY is about creating and tuning Polyglot opening books. For the tuning part it is essential to use engines that understand the opening.

There is no doubt in my mind that the engine that plays the opening best is the Lc0. However Lc0 has a big disadvantage that it can only use one thread because most people have only one GPU installed. This contrary to the AB engines where the analysis can be divided over xx threads, 18 threads in my case with Stockfish 12 as the picture in the link shows.

Now SF12 is a great engine and the NNUE has improved its opening play, but still not good enough to my taste. And so I am looking for alternatives, engines that in your opinion play the opening well and can use multiple threads. Multi threads because analyzing 667K positions (the ProDeo book) is quite time consuming.

There used to be a test that was heavily about openings in my opinion and pretty valid in this area- the 1 hour Kaufman test. Unfortunately I used to test this a much too long time ago, maybe still of limitted interest:

https://www.stmintz.com/ccc/index.php?id=170226

rn1qkb1r/pp2pppp/5n2/3p1b2/3P4/2N1P3/PP3PPP/R1BQKBNR w KQkq - 0 1 id "CCR01"; bm Qb3;
rn1qkb1r/pp2pppp/5n2/3p1b2/3P4/1QN1P3/PP3PPP/R1B1KBNR b KQkq - 1 1 id "CCR02";bm Bc8;
r1bqk2r/ppp2ppp/2n5/4P3/2Bp2n1/5N1P/PP1N1PP1/R2Q1RK1 b kq - 1 10 id "CCR03"; bm Nh6; am Ne5;
r1bqrnk1/pp2bp1p/2p2np1/3p2B1/3P4/2NBPN2/PPQ2PPP/1R3RK1 w - - 1 12 id "CCR04"; bm b4;
rnbqkb1r/ppp1pppp/5n2/8/3PP3/2N5/PP3PPP/R1BQKBNR b KQkq - 3 5 id "CCR05"; bm e5;
rnbq1rk1/pppp1ppp/4pn2/8/1bPP4/P1N5/1PQ1PPPP/R1B1KBNR b KQ - 1 5 id "CCR06"; bm Bcx3+;
r4rk1/3nppbp/bq1p1np1/2pP4/8/2N2NPP/PP2PPB1/R1BQR1K1 b - - 1 12 id "CCR07"; bm Rfb8;
rn1qkb1r/pb1p1ppp/1p2pn2/2p5/2PP4/5NP1/PP2PPBP/RNBQK2R w KQkq c6 1 6 id "CCR08"; bm d5;
r1bq1rk1/1pp2pbp/p1np1np1/3Pp3/2P1P3/2N1BP2/PP4PP/R1NQKB1R b KQ - 1 9 id "CCR09"; bm Nd4;
rnbqr1k1/1p3pbp/p2p1np1/2pP4/4P3/2N5/PP1NBPPP/R1BQ1RK1 w - - 1 11 id "CCR10"; bm a4;
rnbqkb1r/pppp1ppp/5n2/4p3/4PP2/2N5/PPPP2PP/R1BQKBNR b KQkq f3 1 3 id "CCR11"; bm d5;
r1bqk1nr/pppnbppp/3p4/8/2BNP3/8/PPP2PPP/RNBQK2R w KQkq - 2 6 id "CCR12"; bm Bxf7+;
rnbq1b1r/ppp2kpp/3p1n2/8/3PP3/8/PPP2PPP/RNBQKB1R b KQ d3 1 5 id "CCR13"; am Ne4;
rnbqkb1r/pppp1ppp/3n4/8/2BQ4/5N2/PPP2PPP/RNB2RK1 b kq - 1 6 id "CCR14"; am Nxc4;
r2q1rk1/2p1bppp/p2p1n2/1p2P3/4P1b1/1nP1BN2/PP3PPP/RN1QR1K1 w - - 1 12 id "CCR15"; bm exf6;
r1bqkb1r/2pp1ppp/p1n5/1p2p3/3Pn3/1B3N2/PPP2PPP/RNBQ1RK1 b kq - 2 7 id "CCR16"; bm d5;
r2qkbnr/2p2pp1/p1pp4/4p2p/4P1b1/5N1P/PPPP1PP1/RNBQ1RK1 w kq - 1 8 id "CCR17"; am hxg4;
r1bqkb1r/pp3ppp/2np1n2/4p1B1/3NP3/2N5/PPP2PPP/R2QKB1R w KQkq e6 1 7 id "CCR18"; bm Bxf6+;
rn1qk2r/1b2bppp/p2ppn2/1p6/3NP3/1BN5/PPP2PPP/R1BQR1K1 w kq - 5 10 id "CCR19"; am Bxe6;
r1b1kb1r/1pqpnppp/p1n1p3/8/3NP3/2N1B3/PPP1BPPP/R2QK2R w KQkq - 3 8 id "CCR20"; am Ndb5;
r1bqnr2/pp1ppkbp/4N1p1/n3P3/8/2N1B3/PPP2PPP/R2QK2R b KQ - 2 11 id "CCR21"; am Kxe6;
r3kb1r/pp1n1ppp/1q2p3/n2p4/3P1Bb1/2PB1N2/PPQ2PPP/RN2K2R w KQkq - 3 11 id "CCR22"; bm a4;
r1bq1rk1/pppnnppp/4p3/3pP3/1b1P4/2NB3N/PPP2PPP/R1BQK2R w KQ - 3 7 id "CCR23"; bm Bxh7+;
r2qkbnr/ppp1pp1p/3p2p1/3Pn3/4P1b1/2N2N2/PPP2PPP/R1BQKB1R w KQkq - 2 6 id "CCR24"; bm Nxe5;

Admin wrote:

I am working on a new version of POLY 1.7 [ link ]. POLY is about creating and tuning Polyglot opening books. For the tuning part it is essential to use engines that understand the opening.

There is no doubt in my mind that the engine that plays the opening best is the Lc0. However Lc0 has a big disadvantage that it can only use one thread because most people have only one GPU installed. This contrary to the AB engines where the analysis can be divided over xx threads, 18 threads in my case with Stockfish 12 as the picture in the link shows.

Now SF12 is a great engine and the NNUE has improved its opening play, but still not good enough to my taste. And so I am looking for alternatives, engines that in your opinion play the opening well and can use multiple threads. Multi threads because analyzing 667K positions (the ProDeo book) is quite time consuming.

You have a misconception. A GPU is not a CPU. CPU work on cores and threads yes, but has few. But a GPU engine really has many hundreds of cores. That is why they can process huge NN with great speed. So do not lets this bias you, all you are doing is adding many more cores when adding more GPU's. And with RTX Tensor cores as well. Do not think that 2 GPU is equal to 2 threads. 1 GPU can be much better depending on the specs of the GPU.

I don't see my misconception, when I run Lc0 it uses my 1600 GPU for 99%. When I start a second Lc0 run the PC freezes.

Admin wrote:

I don't see my misconception, when I run Lc0 it uses my 1600 GPU for 99%. When I start a second Lc0 run the PC freezes.

Can you explain. Are you trying to run more then 1 Lc0 at the same time on 1 GPU and it freezes? So maybe I do not understand what you are saying.

If this is what you are trying to do. It is possible, but you have to setup Lc0 correctly.

If this is what you are tying to setup. You need to put the 2nd, 3rd.... instance of Lc0 into its own folder. With its own NN. And give it a slightly different name. Just like it is an other engine.
Then you can run many Lc0 on 1 GPU at the same time. And the weird thing I found when doing this, 2 Lc0 running at the same time, run faster then 1/2 speed. And I do not know why.

I think you misunderstood the OP, in more detail - POLY is about book analysis, the ProDeo book contains 677K positions. I analyze them with SF12 at depth=24. Average time per position is about 8 seconds, that means 677,000 * 8 = 1504 hours = 62 days. That is what it would take with Lc0, 62 days. With AB engines that is different, the 677,000 positions are divided over 18 threads and thus the analysis time is limited to 3.5 days.