issue with the avx512 build ?
Minic version 3
Moderator: Ras
-
- Posts: 1872
- Joined: Sat Nov 25, 2017 2:28 pm
- Location: France
Re: Minic version 3
-
- Posts: 2993
- Joined: Wed Mar 08, 2006 10:09 pm
- Location: Germany
- Full name: Werner Schüle
Re: Minic version 3
Yes, normal result should be 10-10
Shall I repeat with the normal skylake compile 1-2 h ?
Shall I repeat with the normal skylake compile 1-2 h ?
Werner
-
- Posts: 1872
- Joined: Sat Nov 25, 2017 2:28 pm
- Location: France
Re: Minic version 3
I think you can go with the skylake build. I removed the avx512 from the repo and try to work on it later. Won't be fixed for 3.16.
Thanks a lot
-
- Posts: 1872
- Joined: Sat Nov 25, 2017 2:28 pm
- Location: France
Re: Minic version 3
ok I had the opporunity to try AVX512 build on a Intel(R) Xeon(R) Silver 4116 CPU @ 2.10GHz compatible with AVX512 instruction set used for NNUE evaluation.
Minic-avx2 is at 530knps while Minicavx512 is around 300knps. For information, in Minic I was using pretty much the same code as the one in openblas for the dot product.
I then tried Seer-avx2 that was at 570knps while Seer-avx512 was a 390knps
It seems using avx512 is maybe decreasing the frequency of the cpu too much and lead to counter-performance in the end.
I have thus removed the avx512 instruction from the repo for the moment, and removed avx512 build of Minic3.16.
Please do not use it if you have already downloaded it.
This will be investigating later again.
Minic-avx2 is at 530knps while Minicavx512 is around 300knps. For information, in Minic I was using pretty much the same code as the one in openblas for the dot product.
I then tried Seer-avx2 that was at 570knps while Seer-avx512 was a 390knps
It seems using avx512 is maybe decreasing the frequency of the cpu too much and lead to counter-performance in the end.
I have thus removed the avx512 instruction from the repo for the moment, and removed avx512 build of Minic3.16.
Please do not use it if you have already downloaded it.
This will be investigating later again.
-
- Posts: 2993
- Joined: Wed Mar 08, 2006 10:09 pm
- Location: Germany
- Full name: Werner Schüle
Re: Minic version 3
Yes, this compile works much better: (40 moves in 2 min)
Minic 3.16 x64 1CPU - Clover 2.4 x64 1CPU : 31,5/52 17-6-29 61% +78
Minic 3.16 x64 1CPU - Clover 2.4 x64 1CPU : 31,5/52 17-6-29 61% +78
Werner
-
- Posts: 1142
- Joined: Thu Dec 28, 2017 4:06 pm
- Location: Argentina
Re: Minic version 3
3.16 says "TCEC Only". Should we use 3.14?
Follow my tournament and some Leela gauntlets live at http://twitch.tv/ccls
-
- Posts: 4718
- Joined: Wed Oct 01, 2008 6:33 am
- Location: Regensburg, Germany
- Full name: Guenther Simon
Re: Minic version 3
That speed diff alone will never explain a 0:20 vs. Clover though...xr_a_y wrote: ↑Tue Oct 12, 2021 8:45 pm ok I had the opporunity to try AVX512 build on a Intel(R) Xeon(R) Silver 4116 CPU @ 2.10GHz compatible with AVX512 instruction set used for NNUE evaluation.
Minic-avx2 is at 530knps while Minicavx512 is around 300knps. For information, in Minic I was using pretty much the same code as the one in openblas for the dot product.
...
@Werner are those games available somewhere for checking?
-
- Posts: 2993
- Joined: Wed Mar 08, 2006 10:09 pm
- Location: Germany
- Full name: Werner Schüle
-
- Posts: 2993
- Joined: Wed Mar 08, 2006 10:09 pm
- Location: Germany
- Full name: Werner Schüle
Re: Minic version 3
Game 1
[pgn] [Event "26b"]
[Site "MSI"]
[Date "2021.10.13"]
[Round "1"]
[White "Minic 3.16AVX512"]
[Black "Clover 2.4 x64 1CPU"]
[Result "0-1"]
[BlackElo "2200"]
[Time "23:22:19"]
[WhiteElo "2200"]
[TimeControl "40/120:40/120:40/120"]
[Termination "normal"]
[PlyCount "115"]
[WhiteType "human"]
[BlackType "program"]
1. Nc3 c5 2. Nf3 e6 3. d4 cxd4 4. Nxd4 a6 5. g3 Qc7 6. Bg2 Nf6 7. O-O Be7
8. a4 Nc6 9. e4 {+0.62/16 3} O-O {-0.04/20 3} 10. Nxc6 {+0.46/16 2} dxc6
{-0.01/24 3} 11. Be3 {+0.54/16 3} a5 {+0.01/24 3} 12. h3 {+1.20/17 2} Rd8
{+0.05/21 3} 13. Qe2 {+1.35/17 3} b6 {+0.01/24 3} 14. Rfd1 {+1.11/18 3}
Rxd1+ {+0.01/25 3} 15. Qxd1 {+1.05/16 2} Ba6 {+0.19/21 3} 16. Qe1 {+1.03/17
2} Rd8 {+0.10/23 3} 17. Kh2 {+1.05/18 2} e5 {+0.05/23 4} 18. f4 {+1.04/16
2} Nd7 {+0.18/24 3} 19. f5 {+1.46/15 3} Bc5 {+0.26/26 3} 20. Bg5 {+1.51/14
3} Nf6 {+0.47/25 3} 21. Bxf6 {+1.12/16 3} gxf6 {+1.04/26 3} 22. h4
{+0.21/20 13} Bb4 {+1.19/24 3} 23. Bf3 {+0.26/19 12} Rd4 {+1.23/25 3} 24.
Qe3 {+0.35/21 2} Qd8 {+1.64/26 4} 25. Rg1 {-0.02/21 8} Rd2+ {+1.71/24 3}
26. Rg2 {-0.22/23 7} Rxg2+ {+1.21/26 3} 27. Kxg2 {-0.90/19 1} Bc5 {+1.79/25
3} 28. Qc1 {+1.87/11} Qd4 {+1.95/23 3} 29. Qe1 {-0.48/19 1} Kg7 {+1.63/25
4} 30. g4 {-0.92/18} h6 {+2.09/24 3} 31. Be2 {-2.41/19 2} Bxe2 {+2.10/26 3}
32. Qxe2 {-2.23/23 1} Qg1+ {+2.03/25 3} 33. Kh3 {-2.46/20 3} Bf2 {+2.39/28
3} 34. Qf3 {-1.65/19 1} Qf1+ {+2.59/30 3} 35. Qg2 {-2.51/22} Qe1 {+2.87/28
3} 36. Qf3 {-2.80/26 1} Bxh4 {+3.04/26 3} 37. b3 {-1.95/24 1} Bf2 {+2.99/27
3} 38. Kg2 {-1.69/18} Qg1+ {+3.21/26 3} 39. Kh3 {-3.59/20 2} Qf1+ {+3.30/30
3} 40. Qg2 {-4.40/21 2} Qe1 {+3.19/27 3} 41. Qf3 {-4.55/26 10} Qf1+
{+3.06/27 3} 42. Qg2 {-3.94/21} Qe1 {+3.23/27 3} 43. Qf3 {-3.96/27 12} Kf8
{+3.13/27 3} 44. Nd1 {-4.88/26 15} Bd4 {+3.98/24 3} 45. Qd3 {-3.98/16} Kg7
{+4.05/23 2} 46. Kg2 {-5.03/20 2} b5 {+3.88/24 2} 47. axb5 {-6.27/23 17}
cxb5 {+5.54/21 3} 48. c3 {-7.43/19 1} Ba7 {+6.26/24 3} 49. Qc2 {-8.38/22
12} Qg1+ {+7.21/25 4} 50. Kh3 {-11.27/22 9} Qf1+ {+8.22/23 4} 51. Kg3
{-11.57/16} Qf4+ {+8.56/23 2} 52. Kg2 {-6.18/12} Qxg4+ {+9.36/23 2} 53. Kf1
{-8.35/13} Qg1+ {+14.25/24 4} 54. Ke2 {-4.47/12} h5 {+16.09/23 9} 55. Kd2
{-8.06/17 1} Qg5+ {+19.40/25 5} 56. Ke2 {-13.29/23 6} h4 {+20.93/24 6} 57.
Qd2 {-13.05/19 1} Qg2+ {+23.67/21 2} 58. Kd3 {-14.31/18 Arena
Entscheidung} 0-1 [/pgn]
I get diffent evals when using AVX2 compile. e.g. move 24 Qe3 -0.94/21 instead of + 0.35/21
actual result with avx2 Minic 3.16 x64 1CPU - Clover 2.4 x64 1CPU : 26,0/46 14-8-24 57% +49
[pgn] [Event "26b"]
[Site "MSI"]
[Date "2021.10.13"]
[Round "1"]
[White "Minic 3.16AVX512"]
[Black "Clover 2.4 x64 1CPU"]
[Result "0-1"]
[BlackElo "2200"]
[Time "23:22:19"]
[WhiteElo "2200"]
[TimeControl "40/120:40/120:40/120"]
[Termination "normal"]
[PlyCount "115"]
[WhiteType "human"]
[BlackType "program"]
1. Nc3 c5 2. Nf3 e6 3. d4 cxd4 4. Nxd4 a6 5. g3 Qc7 6. Bg2 Nf6 7. O-O Be7
8. a4 Nc6 9. e4 {+0.62/16 3} O-O {-0.04/20 3} 10. Nxc6 {+0.46/16 2} dxc6
{-0.01/24 3} 11. Be3 {+0.54/16 3} a5 {+0.01/24 3} 12. h3 {+1.20/17 2} Rd8
{+0.05/21 3} 13. Qe2 {+1.35/17 3} b6 {+0.01/24 3} 14. Rfd1 {+1.11/18 3}
Rxd1+ {+0.01/25 3} 15. Qxd1 {+1.05/16 2} Ba6 {+0.19/21 3} 16. Qe1 {+1.03/17
2} Rd8 {+0.10/23 3} 17. Kh2 {+1.05/18 2} e5 {+0.05/23 4} 18. f4 {+1.04/16
2} Nd7 {+0.18/24 3} 19. f5 {+1.46/15 3} Bc5 {+0.26/26 3} 20. Bg5 {+1.51/14
3} Nf6 {+0.47/25 3} 21. Bxf6 {+1.12/16 3} gxf6 {+1.04/26 3} 22. h4
{+0.21/20 13} Bb4 {+1.19/24 3} 23. Bf3 {+0.26/19 12} Rd4 {+1.23/25 3} 24.
Qe3 {+0.35/21 2} Qd8 {+1.64/26 4} 25. Rg1 {-0.02/21 8} Rd2+ {+1.71/24 3}
26. Rg2 {-0.22/23 7} Rxg2+ {+1.21/26 3} 27. Kxg2 {-0.90/19 1} Bc5 {+1.79/25
3} 28. Qc1 {+1.87/11} Qd4 {+1.95/23 3} 29. Qe1 {-0.48/19 1} Kg7 {+1.63/25
4} 30. g4 {-0.92/18} h6 {+2.09/24 3} 31. Be2 {-2.41/19 2} Bxe2 {+2.10/26 3}
32. Qxe2 {-2.23/23 1} Qg1+ {+2.03/25 3} 33. Kh3 {-2.46/20 3} Bf2 {+2.39/28
3} 34. Qf3 {-1.65/19 1} Qf1+ {+2.59/30 3} 35. Qg2 {-2.51/22} Qe1 {+2.87/28
3} 36. Qf3 {-2.80/26 1} Bxh4 {+3.04/26 3} 37. b3 {-1.95/24 1} Bf2 {+2.99/27
3} 38. Kg2 {-1.69/18} Qg1+ {+3.21/26 3} 39. Kh3 {-3.59/20 2} Qf1+ {+3.30/30
3} 40. Qg2 {-4.40/21 2} Qe1 {+3.19/27 3} 41. Qf3 {-4.55/26 10} Qf1+
{+3.06/27 3} 42. Qg2 {-3.94/21} Qe1 {+3.23/27 3} 43. Qf3 {-3.96/27 12} Kf8
{+3.13/27 3} 44. Nd1 {-4.88/26 15} Bd4 {+3.98/24 3} 45. Qd3 {-3.98/16} Kg7
{+4.05/23 2} 46. Kg2 {-5.03/20 2} b5 {+3.88/24 2} 47. axb5 {-6.27/23 17}
cxb5 {+5.54/21 3} 48. c3 {-7.43/19 1} Ba7 {+6.26/24 3} 49. Qc2 {-8.38/22
12} Qg1+ {+7.21/25 4} 50. Kh3 {-11.27/22 9} Qf1+ {+8.22/23 4} 51. Kg3
{-11.57/16} Qf4+ {+8.56/23 2} 52. Kg2 {-6.18/12} Qxg4+ {+9.36/23 2} 53. Kf1
{-8.35/13} Qg1+ {+14.25/24 4} 54. Ke2 {-4.47/12} h5 {+16.09/23 9} 55. Kd2
{-8.06/17 1} Qg5+ {+19.40/25 5} 56. Ke2 {-13.29/23 6} h4 {+20.93/24 6} 57.
Qd2 {-13.05/19 1} Qg2+ {+23.67/21 2} 58. Kd3 {-14.31/18 Arena
Entscheidung} 0-1 [/pgn]
I get diffent evals when using AVX2 compile. e.g. move 24 Qe3 -0.94/21 instead of + 0.35/21
actual result with avx2 Minic 3.16 x64 1CPU - Clover 2.4 x64 1CPU : 26,0/46 14-8-24 57% +49
Werner
-
- Posts: 1872
- Joined: Sat Nov 25, 2017 2:28 pm
- Location: France
Re: Minic version 3
3.16 can be used. I will fix the release note.