Author: Tom Kerrigan
Date: 08:38:21 04/08/00
Go up one level in this thread
On April 08, 2000 at 09:35:39, Robert Hyatt wrote: >On April 08, 2000 at 01:34:50, Tom Kerrigan wrote: > >>On April 07, 2000 at 23:20:54, Robert Hyatt wrote: >> >>>>I still disagree. Didn't DB do a quick evaluation in ~4 cycles? So going back to >>>>my example, the time to search 2/5 capture moves using MVV/LVA is: >>>> >>>>generate moves = 2 * 2 >>>>make moves = 1 * 2 >>>>eval = 2 * 8 >>>>=> 22 cycles >>>> >>>>And if you use SEE: >>>> >>>>generate moves = 2 * 5 >>>>SEE = 5 * 5 >>>I don't see 5*5 cycles in SEE. I see 5 to enumerate the capturing pieces, >>>5 to minimax the score. 10 total... That is definitely cheaper than searching >>>a node, which takes 10 clock cycles in DB. Because in the above case you >>>would have to search up to 10 nodes (5 captures per side) which is 100 clocks. >> >>I meant that the SEE would take on average 5 cycles, and it would have to be >>done for 5 different moves. >> >>Searching a node might take 10 cycles for DB. And adding a SEE would add a >>significant number of cycles to many of those nodes. That's exactly my point. >> >>-Tom > > >Then we strayed... my original comment was "the ultimate SEE would be too slow >_unless_ it was designed as a special-purpose piece of hardware." I have no >doubt I could design a device to sit on the bus somewhere and handle SEE >analysis, and do it inside the cycle time required to handle a single normal >software SEE. Probably at least 100 times faster, if not more... Oh, whoops. I didn't realize that's what the original conversation was about. -Tom > >and there is another point. SEE can be used to reduce the overall size of >the tree by 50%, easily. Which would further widen the gap. In hardware that >does _everything_ it might well be a different issue... Or the hardware will >work differently than you expect... IE the find-victim cycle is _not_ needed >in SEE. that goes out. The find-aggressor cycle could, in one cycle, enumerate >all the pieces that attack the target square. So in one cycle, we enumerate all >the black and white pieces that attack the square. It _might_ require 5 cycles >to do the swap calculations although I doubt it. I believe that the typical >SEE could be done in around 5 cycles which is not horribly slow...
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.