Steve Maughan wrote:The time control probably has a lot to do with the results. At this fast time control you're not super-saturating the smaller hash table, so the benefits of the larger hash table are not shining through.
Steve
I went to the safe side, well saturating the hash table, and measured another way the benefit of hash table of Houdini 4, with time-to-depth (TTD) on 100 positions. Average time per position 30 seconds, speed on 4 cores about 8-9Mnps, the hash table needed is well above 1GB. I used 4MB and 256MB hash tables for comparison on TTD, well inside the necessary:
Code: Select all
H4 on 4 cores:
100 positions Depth=21 4MB time: 53:02
100 positions Depth=21 256MB time: 32:32
The benefit, assuming 100*log[time] Elo points per doubling time is
11 Elo points from doubling the hash table, when the hash table is clearly saturated.
A similar result I get in direct games with SF, but direct games are more tedious: 1' + 1'' hash 1MB and hash 16MB, both well saturated in games:
Code: Select all
Score of SF 16MB vs SF 1MB: 293 - 96 - 834 [0.58] 1223
ELO difference: 56
Finished match
Here 14 Elo points for doubling hash table, but at somewhat shorter TC. I guess 11 Elo points is safe to say at blitz and a bit longer TC for SF too.