I've been trying BBC online (and have version 1.1 and 1.2 installed as well). Some observations:
- In the Fritz GUI, BBC only reports 650 nodes (not kNodes... 650, as in under 1000) per second. I assume you're passing this wrong to the GUI. (I haven't checked the code). This would be extremely slow.
- Online, BBC 1.3 reports 150.xxx kNodes. That would be 150-250 kNodes/second. That seems more correct, but it's still slow; it's 1/4th to half of the speed of my three year old midrange phone, or half the speed of a Raspberry Pi 4. Where is the engine running?
- The NNUE 1.3 version is much weaker than the BBC-version I played earlier. That version was quite difficult to defeat for me even at instant response. (I'm a decent, but not super-strong player; somewhere between 1800 and 2000 FIDE Elo depending on how much effort I put into a game.) For earlier versions of BBC I would actually need to put in some effort to not lose a game. Here's a game I played against 1.3 + NNUE, where my response time was basically as fast as you'd see in a bullet or 3 minute blitz game... (and I suspect I made many mistakes and left a lot of chances unused). Basically, I put no effort at all into this game. Especially black's 36 ... h5 boggles my mind. It just completely ignores an attacked bishop. During the entire game the engine didn't even make an effort to get castled.
Well, what can I say:
1. Knps calculation was wrong in early versions due to the bugs which I've fixed later
2. Current Knps of 150.xxx is 65% compared to handcrafted eval. NNUE slows down performance significantly, especially bearing in mind the fact that I didn't involve incremental updates optimization for simplicity
3. And regarding strength, well, I guess NNUE just needs a bit more time to search and instant response is not an option, I will probably remove it.
I've been trying BBC online (and have version 1.1 and 1.2 installed as well). Some observations:
- In the Fritz GUI, BBC only reports 650 nodes (not kNodes... 650, as in under 1000) per second. I assume you're passing this wrong to the GUI. (I haven't checked the code). This would be extremely slow.
- Online, BBC 1.3 reports 150.xxx kNodes. That would be 150-250 kNodes/second. That seems more correct, but it's still slow; it's 1/4th to half of the speed of my three year old midrange phone, or half the speed of a Raspberry Pi 4. Where is the engine running?
- The NNUE 1.3 version is much weaker than the BBC-version I played earlier. That version was quite difficult to defeat for me even at instant response. (I'm a decent, but not super-strong player; somewhere between 1800 and 2000 FIDE Elo depending on how much effort I put into a game.) For earlier versions of BBC I would actually need to put in some effort to not lose a game. Here's a game I played against 1.3 + NNUE, where my response time was basically as fast as you'd see in a bullet or 3 minute blitz game... (and I suspect I made many mistakes and left a lot of chances unused). Basically, I put no effort at all into this game. Especially black's 36 ... h5 boggles my mind. It just completely ignores an attacked bishop. During the entire game the engine didn't even make an effort to get castled.
I've just fixed horrible bug...
The way it was working (and even crushing VICE with this bug!) is the following: on obtaining a threshold material phase eval was changing from NNUE to handcrafted BUT there was an issue of score scaling - I didn't do any scaling so when nnue evaluated position like +4.25 it could still be the equal material, but only positional advantage (due to NNUE) and now imagine it drops to my handcrafted eval - it starts thinking that it's material up and starting giving pieces away, thinking it's much better)
So I've updated NN to latest (they've switched to new one in SF yesterday) and got rid of handcrafted eval as well.
Btw new NNUE performs much better in endgame.
I've already updated code in online version.
Now after some tests I'll update files in release as well.