What is the general viewpoint on table memory allocation
Shall i stick with default value (128 Mb) for my laptop (Lenova G70-80)
I tend to use 256 Mb for my desktop computer though.
The last few months i've been in doubt if allocating 256 Mb for an 2-core laptop can negatively impact overall performance, regardless of time control.
Nordlandia wrote:What is the general viewpoint on table memory allocation
Shall i stick with default value (128 Mb) for my laptop (Lenova G70-80)
I tend to use 256 Mb for my desktop computer though.
The last few months i've been in doubt if allocating 256 Mb for an 2-core laptop can negatively impact overall performance, regardless of time control.
2-core fits 128 Mb or above
8-core fits 256 Mb or above
The best settings will depend on time controls you use and what size caches, Threads, and NUMA nodes are in a specific machine. This varies a lot. The point of Table Memory is to save time, by storing evaluation information in memory instead of recalculating it each time Komodo evaluates a position. My suggestion is first set Hash to an appropriate size using the instructions in setHash.txt that comes with Komodo. Then figure out the approximate time Komodo will analyze a position using whatever time controls you are using. Lets say for 40 moves in 40 minutes is the time control you will use, then 4/40 = 1 minute on average per move. Analyze a position using say 128 Mb or Table memory and not the nodes per second the GUI reports. Reload Komodo, change the Table Memory to a bigger or smaller value, then analyze again for the same time. Pick whatever setting gives you the most nodes per second.
I would not expect to see a big nps change unless you are using a very fast time control, where most of the Table Memory can fit in the device cache. For very long analysis you can use a bigger Table Memory setting. In general use, the default will be fine for most people, but we allow up to 1024 MB of Table Memory for people with huge, many Thread machines.
SF has made virtually no progress rating-wise since the release of SF 8. What have been making progress are various derivatives of SF such as Brainfish.
On Pohl's site, SF progress is +16 Elo from SF 8, and Asmfish is +17 Elo over the same time span.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
Regression test showed +9 elo in 3 month I think.
But now Stockfish dev is probably around +15 elo.
It is the first slow down of progress since long time for Stockfish, but that might change soon.
So in the meantime, Komodo can catch up a little Stockfish.
