Author: Robert Hyatt
Date: 22:01:41 03/29/03
Go up one level in this thread
On March 29, 2003 at 07:45:32, K. Burcham wrote: > > >I think also that the ram available for resources and large hash has exceeded >the mhz rating for now. In other words with four slots of 256, or two slots of >512, there is more than enough memory for our hash settings for now. Here's some rough math. I'm running tests to see how this holds up in practice and will post the results tomorrow. I'm running the first 6 bt2630 positions to a fixed depth, with hash sizes starting at 48K bytes and doubling each run. I'll post the sizes and the total search time for the 6 positions. Back to the math: Crafty searches about 2.5M nodes per second as an upper limit in middlegame positions, using my dual 2.8ghz xeon. For a search in a 40 moves/2hrs game, Crafty will average roughly 6 minutes a move, counting the first N moves in zero time from book, and the time saved by pondering. 6 minutes = 360 seconds = 2.5 * 360 * M = 900M positions. Some fraction of that is not stored in Crafty, Until I get the actual data out, I'll guess that at least 3/4 of those positions are not hashed. Meaning I need to store 225M positions (maybe). At 16 bytes per position, 3.6 gigabytes of RAM are needed to hold exactly that many entries. Since the hash signatures are _not_ uniformly distributed, I'd probably go for a hash table 2x that big to have some confidence that not many positions are getting replaced which hurts tree size. 8 gigs roughly. Pretty big. If you assume 90% of searched nodes are at depth=0 and beyond, that means 10% or 90M positions need to be stored, or 1.4 gigs of RAM, and to be safe, double that to 3 gigs. I'm not sure about the .75 or .90 multiplier for the q-search. And it is also not easy to prove that storing less than the full tree is sub-optimal, although logically it should be easy to see. My results should be ready tomorrow morning and I'll post them as soon as I can... > >If the processors were to take a big jump in size, then the larger hash would be >helpful. If we had a 6000 mhz processor, with 3 gigs of ram and hash set at 1000 >megs, then I think this hash size can be used. It seems to me that over the last >ten years, the total amount of memory that has become available in todays >systems has exceeded the processing power. > >I think someday we will look back on 1000 megs of hash as small. > >kburcham I think that one day folks will look back and see 1000 terabytes as small.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.