TalkChess.com

[Daniel Shawul]

A ton of very serious bugs were fixed. For starters the uct select was completely broken producing rather strange shallow and uniform tree. But after the many bug fixes, it starts producing good result as shown below. Also the occupancy of my last test was at 50% so I have a chance to increase that to 100% by simply changing configurations. Infact the occupancy has been slightly increased on my current gpu to 67% after the register usage reduced when the bugs are fixed. So that directly translated to performance increase. On capability >1.3 block size of 256 gives 100% occupancy so that will be doubled performance if everything goes well.
One other thing is that now there is absolutely no synchronization (__syncthreads) except for the inherent warp sync. A warp is the smallest working unit to fetch a node from DRAM; previously it was a block. The threads in a warp can diverge since each does its own monte carlo game on the same initial board. That has speeded up the code and simplified it as well.
On the initial position e4! (this is HEX btw!) is chosen as the best move and it was expanded to depth 7 after 45 mil simulation. Time is decreased to 5.375sec compared to previous ~5.5. Tree is also very selective.