|
|
Post by matejst on Apr 29, 2022 14:38:19 GMT
Ed Schroeder published Rebel 15, with solid improvements. Unfortunately, no SSE versions. While we all know that "NNUE evaluation strongly relies on the AVX2 instruction set and there might come a moment older computers become an obstacle to progress", Stockfish, Komodo, Koivisto, Berserk, Seer, Rubichess all have SSE versions. Go figure.
|
|
|
|
Post by Ozymandias on Apr 29, 2022 16:01:52 GMT
Is this the one you were waiting for?
What a bummer. Not even vanilla x64? Anyone with a CPU older than 6 years is out, right?
|
|
|
|
Post by matejst on Apr 29, 2022 21:07:33 GMT
Yep. That's it. Seems that it is a question of architecture.
Edit: I am not a programmer, but... The Toga IV engine was developed and written before AVX2. The code for the NN was added by Chris. Although a ignoramus in these affairs, it seems to me that compiling for sse4.1 avx should be just a thing of... compiling. We see that SF has several versions daily! The maintenance of the code cannot be that complicated. I feel it is more a question of good will than a technical question. But... what do I know.
Edit 2: When I think that I am probably the only true user of Rebel (the only one using this engine as his main/only engine), it was very clever not to even try to make sse builds. But this reveals mainly my stupidity. I will switch now to another engine. Eman perhaps? Another one?  |
|
|
|
Post by Ozymandias on May 4, 2022 18:51:31 GMT
|
|
|
|
Post by matejst on May 4, 2022 22:51:27 GMT
The NN is all right, the search is atrocious. The previous version misses mates in 6 because it does not have check extensions. Furthermore, the evaluation is aggressive, but when the engine cannot find technically winning moves...
Ed started with a good engine - Marvin - but then he had the awful idea to switch to Fruit and Toga, engines he simply does not understand.
When analysing for OTB play, I really don't need SF, and a fast engine makes things difficult to understand. Imagine a complicated position, which is a draw at dept 40. A slow engine, that gives me an advantage at depth 20 is more useful to me. With multiPV, it will show me a lot of bad moves and how to punish them. It will also show which positions to avoid.
A good engine for Otb preparation was Orion 0.8, but no multiPV. Slow, tactically very good, eval based on SF, very good. Fire with one of Dkappe nets was ok too. Dragon 2 also, but I found the Benjamin NN really dynamic and interesting. Unfortunately, Ed was about experimenting instead of fixing the few weaknesses that posed problems with the first NNs (bishops of diff colors, bad eval scaling, a few other simple things).
|
|
|
|
Post by matejst on May 14, 2022 8:14:16 GMT
Ed managed to find the bugs and made a SSE3 version.
|
|
|
|
Post by matejst on May 17, 2022 20:00:19 GMT
A new Rebel 15x2 was published today. The net is twice bigger. A bit stronger than Rebel 15.
|
|
|
|
Post by matejst on Jul 13, 2022 13:52:24 GMT
New Rebel 15.1. Much faster (2x+), with quality evaluation. Most of the job has been done by modifying the structure of the NN. So far, good impressions.
|
|
|
|
Post by Ozymandias on Jul 13, 2022 20:24:40 GMT
Yes, I noticed you didn't like the post on the original forum.  |
|
|
|
Post by matejst on Jul 13, 2022 21:08:40 GMT
Yes, I noticed you didn't like the post on the original forum. I did not see it. But in general, I like what has been done so far. |
|
|
|
Post by Ozymandias on Dec 26, 2022 7:34:27 GMT
When analysing for OTB play, I really don't need SF On the topic of SF. I've made several test with some versions since August 17 (as per Stefan Pohl's recommendation) all the way to December 21, including 15.1. It's safe to say there's been no progress in the SF front for over 4 months now. Dragon is by its own admission out of the race. Very soon, those strange engines you like might catch up.  |
|
|
|
Post by matejst on Dec 26, 2022 8:52:14 GMT
I don't think so. SF still has not only the fastest search, but its nets are the best, since they are based more and more on Leela's data. I tried Fat Titz with its 150+ mb net, it seems ok, and Norman Schmidt version (Big Ballz) seems a bit better, but it's just a derivative of SF.
I don't understand why not more authors use the freely available Lc0 data, why only the SF team. Imho, a Lc0 distilled for CPU could be my choice.
Stylistically, there's no differences whatsoever. The first Rebel, based on Benjamin, was a great engine, but it has been "dulled" since then. Wasp remains a bit more aggressive, but real in simple positions. I tried Minic lately, and I like what Vivien does. Berserk is also progressing nicely. Overall, it makes no difference whatsoever when I am analyzing.
A word about Dragon. I don't understand the fuss for 15 Elos added in the last three versions.
BTW, these days I play against the personalities in ChessMaster. I feel sometimes quite stupid, because I spend minutes thinking about a position, possible complications, and my opponent blunders in the simplest possible way. But I feel I have progressed nicely. Positionally, I play better. Still a bit wobbly when calculating.
|
|
|
|
Post by Ozymandias on Dec 26, 2022 9:35:15 GMT
I don't think so. SF still has not only the fastest search, but its nets are the best, since they are based more and more on Leela's data. [...] A word about Dragon. I don't understand the fuss for 15 Elos added in the last three versions. Net progress is even worse. If you look at their page, of the last 250 nets marked green, only four actually made it into the engine, and the last one (more heavily dependent on Lc0 data than ever before) shows a 1.2 Elo increase in my test. Search improvements are also stagnating. Long gone are the days of the yearly 100 Elo improvements of the 90's and early 00's. Gone are the days of the yearly 50 Elo improvements (excluding Rybka and Houdini releases). Even gone are now the 20 Elo yearly improvements. We'll probably be looking now at a 10 Elo/year progress.
I suspect many people don't which is why 3.2 could be the last in the series. |
|
|
|
Post by matejst on Dec 26, 2022 16:13:42 GMT
I downloaded this last dev. version almost entirely based on Lc0 data. The problem with SF was that it was always an optimisation of known ideas, and, while an engineering feat, it gave nothing new, revolutionary to engine development. Seer, Winter, SlowChess achieved much more with less, trying new stuff. The SF team should experiment more. Ditto for Komodo.
|
|
|
|
Post by Ozymandias on Dec 26, 2022 16:20:07 GMT
I downloaded this last dev. version almost entirely based on Lc0 data. The problem with SF was Is it still? |
|
