Poker-AI.org

Recently I've read Amax' post in the old forum, where he said, that his CS-CFRM algorithm runs at 550k iterations per second.

I have been using the pure cfr open source code and even when I completelly diasble bucketing and hand evaluation, which means I just always return 0 (for testing purposes, of course this gives wrong results) I only manage to get about 30k iterations per seconds, even with multithreading. So even just the recursion through the tree takes way longer.

Is the algorithm just inefficient? What can I do to speed it up?

is this for nl or fl?
if nl, what betting abstraction & stack size?

This was for FL.

you might be encountering false sharing by making every thread edit that single location in memory. if you get rid of the writes you might see a drastic performance increase. or you can actually use bucketing.