TalkChess.com

"Hi
forceNNUE=true
will result in a full NNUE eval, and thus a slow down.
I highly recommand setting this too false to use "hybrid" evaluation (threshold is 600cp currently)."


Thanks Vivien,
settings changed here.

Werner wrote: ↑Thu Oct 15, 2020 1:36 pm "Hi
forceNNUE=true
will result in a full NNUE eval, and thus a slow down.
I highly recommand setting this too false to use "hybrid" evaluation (threshold is 600cp currently)."


Thanks Vivien,
settings changed here.

If I understand correctly:

1. Without setting an NNUE file evaluation is always traditional.
2. With an NNUE file but Force NNUE off the evaluation is traditional below 600 cp but NNUE is used above 600 cp.*
3. With an NNUE file and Force NNUE on the evaluation is always NNUE.

*What if switching between evaluation methods the 600 cp threshold also switches?

Gabor Szots wrote: ↑Thu Oct 15, 2020 3:06 pm

Werner wrote: ↑Thu Oct 15, 2020 1:36 pm "Hi
forceNNUE=true
will result in a full NNUE eval, and thus a slow down.
I highly recommand setting this too false to use "hybrid" evaluation (threshold is 600cp currently)."


Thanks Vivien,
settings changed here.

If I understand correctly:

1. Without setting an NNUE file evaluation is always traditional.
2. With an NNUE file but Force NNUE off the evaluation is traditional below 600 cp but NNUE is used above 600 cp.*
3. With an NNUE file and Force NNUE on the evaluation is always NNUE.

*What if switching between evaluation methods the 600 cp threshold also switches?

1. correct
2. no. It is the other way around: NNUE if material imbalance < 600 and standard if not (for already decided positions)
3. correct

xr_a_y wrote: ↑Thu Oct 15, 2020 3:09 pm1. correct
2. no. It is the other way around: NNUE if material imbalance < 600 and standard if not (for already decided positions)
3. correct

Thanks, it was not clear that the difference was material.

Hi, thanks for reporting. I am pretty sure that sure work. Something relative to path may be wrong. Easiest way to be sure is to put net in the same directory and just use its name as input.

2.48 has less logging about NNUE, I am not sure it loaded net in your example. Check if nps are small to be sure the net is loaded under 2.48.
[/quote]



Putting the net in the same directory and using only the name as input made it work.

Thank you.

Hm,
cannot get running minic_2.52_mingw_x64_skylake.exe in a GUI (Arena says 0MB; in Shredder GUI crashes).
works ok im command promt.
info string This is Minic version 2.52 (NNUE available)
info string Info 2020-10-16 18:24:52-750: No NNUE net given
info string Info 2020-10-16 18:24:52-752: Classical evaluation enabled.
setoption name NNUEFile value nascent_nutrient.bin
info string Info 2020-10-16 18:26:03-462: NNUE evaluation using nascent_nutrient.bin enabled.
go depth 2
info depth 1 score cp 1 time 1 nodes 23 nps 22999 seldepth 1 pv d2d4 tbhits 0
info depth 2 score cp 143 time 1 nodes 45 nps 44999 seldepth 2 pv d2d4 b8c6 tbhits 0
bestmove d2d4 ponder b8c6:
reason here: it happens, when I use 256 MB hash instead of 128 MB default. Here Arena is ok, Shredder GUI still crashes.

Ok let me check that on my Windows box.
If 2.51 was working this is due to my memory leak patch. So something wrong with aligned memory under Windows...

xr_a_y wrote: ↑Fri Oct 16, 2020 6:31 pm Ok let me check that on my Windows box.
If 2.51 was working this is due to my memory leak patch. So something wrong with aligned memory under Windows...

As usual, I pushed too fast . Fixed in 2.53, will release tonight. Very sorry. Thanks for reporting.

xr_a_y wrote: ↑Fri Oct 16, 2020 7:17 pm As usual, I pushed too fast . Fixed in 2.53, will release tonight. Very sorry. Thanks for reporting.

Thanks Vivien,
2.53 works now ok under Windows. Do you think there is an Elo difference to 2.51 - or it only causes somtimes crashes?

Werner wrote: ↑Sat Oct 17, 2020 10:08 am

xr_a_y wrote: ↑Fri Oct 16, 2020 7:17 pm As usual, I pushed too fast . Fixed in 2.53, will release tonight. Very sorry. Thanks for reporting.

Thanks Vivien,
2.53 works now ok under Windows. Do you think there is an Elo difference to 2.51 - or it only causes somtimes crashes?

No Elo diff versus 2.51 (just memory leak fixes).