That is ofc not true. 1650 is on par with 1070Ti, i.e. in best case around 7k nps from the start position. 2060 is over 21k nps from the start position. So difference is 3x at least.
And there is an additional caveat. 1650 achieves this performance in full FP16, meaning FP16 accumulation in addition to FP16 multiplication.
RTX architectures perform FP16 multiplication too, but FP32 accumulation, so even with the same Lc0 running on RTX should show a bit better performance (due to more accurate inference).
jp wrote: ↑Wed Apr 15, 2020 7:45 am
Well, that's because (we believe) chess is a draw. You don't know that you haven't missed wins when your opponents made mistakes that ended as draws.
I know, I'm not blind. What I can say about that is that the wins I've missed would have been missed with better hardware, and would have been found with my hardware with better analysis methods. I failed to win because I became lazy and thought the games were already decided, so I played considerably below my level. My usual methods would have sufficed to win them if I used them...
zullil wrote: ↑Wed Apr 15, 2020 3:10 pmOK, here's an easy one for you:
[d]8/1p2B3/4k3/7p/p1Pp3P/P2P1K2/8/4b3 w - - 6 51
What move(s) are "best"?
I find suspect that you post a position where I have no problem reaching depth 60 very quick with my hardware, in discussions where a position where someone could outsearch me is called for. But now I'm really curious to see how you tell me Bg5 (found in 15 minutes) isn't the "best" move.
zullil wrote: ↑Wed Apr 15, 2020 3:10 pmOK, here's an easy one for you:
[d]8/1p2B3/4k3/7p/p1Pp3P/P2P1K2/8/4b3 w - - 6 51
What move(s) are "best"?
I find suspect that you post a position where I have no problem reaching depth 60 very quick with my hardware, in discussions where a position where someone could outsearch me is called for. But now I'm really curious to see how you tell me Bg5 (found in 15 minutes) isn't the "best" move.
It mainly depends on the developing of NVIDIA GPUs.
zullil wrote: ↑Wed Apr 15, 2020 3:10 pmOK, here's an easy one for you:
[d]8/1p2B3/4k3/7p/p1Pp3P/P2P1K2/8/4b3 w - - 6 51
What move(s) are "best"?
I find suspect that you post a position where I have no problem reaching depth 60 very quick with my hardware, in discussions where a position where someone could outsearch me is called for. But now I'm really curious to see how you tell me Bg5 (found in 15 minutes) isn't the "best" move.
OK, I think you just lost the game, because your hardware/method failed you.
Perhaps someone with better hardware than mine can confirm that Ke2 is the only move here that holds the draw?
Alayan wrote: ↑Thu Apr 16, 2020 1:58 pm
This is Stockfish-Komodo with black and white reversed, isn't it ?
Then Bg5 is guaranteed to lose, and Ke2 likely holds (or at least, we don't know how to beat it).
Yes, it is. The original position was discussed at this forum, so I intentionally obfuscated the position so that searching the FEN would not be immediately useful. Here's what I currently see after Bg5:
Alayan wrote: ↑Thu Apr 16, 2020 1:58 pm
This is Stockfish-Komodo with black and white reversed, isn't it ?
Then Bg5 is guaranteed to lose, and Ke2 likely holds (or at least, we don't know how to beat it).
Yes, it is. The original position was discussed at this forum, so I intentionally obfuscated the position so that searching the FEN would not be immediately useful. Here's what I currently see after Bg5:
That's after a deep Stockfish search of the original position with MultiPV = 2, followed by forward/backward analysis of the Bg5 line.
20 threads, 64 GB hash table.
I really don't know what are you taking about. SFdev, 6-men, MultiPV=2, depth 80, Bg5 and Kd2 are a transposition so lines are identical, score -1.07.
Line is 51. Bg5 Ba5 52. Ke2 Bc7 or 51.Ke2 Ba5 52. Bg5 Bc7
Alayan wrote: ↑Thu Apr 16, 2020 1:58 pm
This is Stockfish-Komodo with black and white reversed, isn't it ?
Then Bg5 is guaranteed to lose, and Ke2 likely holds (or at least, we don't know how to beat it).
Yes, it is. The original position was discussed at this forum, so I intentionally obfuscated the position so that searching the FEN would not be immediately useful. Here's what I currently see after Bg5:
That's after a deep Stockfish search of the original position with MultiPV = 2, followed by forward/backward analysis of the Bg5 line.
20 threads, 64 GB hash table.
I really don't know what are you taking about. SFdev, 6-men, MultiPV=2, depth 80, Bg5 and Kd2 are a transposition so lines are identical, score -1.07.
Line is 51. Bg5 Ba5 52. Ke2 Bc7 or 51.Ke2 Ba5 52. Bg5 Bc7
Please let us know if you discover that your initial analysis is incorrect.
