Author: Eugene Nalimov
Date: 09:54:35 06/21/02
Go up one level in this thread
Sorry, pressed the wrong button. Of course if cache is 4Mb and hash table is only 2Mb than we'll not go to the main memory. But larger hash table will buy you much more, even if access will be much slower. Yes, newer types of RAM give you some speedup, but they help bandwidth much more than latency. And in the case of hash table probe you are interested in latency. Let's take a look at the timings of one of the modern high-end CPUs I am familiar with: L1 read 1 CPU cycle L2 read 5 CPU cycles L3 read 12 CPU cycles Main memory read 180 CPU cycles CPU is server one, so it is clocked not high by the high end desktop standards. For 2.5GHz CPU main memory access will be slower (in CPU cycles) regardless of the used memory type. Eugene On June 21, 2002 at 12:46:02, Eugene Nalimov wrote: >On June 21, 2002 at 02:02:58, Dann Corbit wrote: > >>On June 21, 2002 at 01:39:11, Eugene Nalimov wrote: >>[snip] >>>And each hash table probe (main memory access) costs several hundred CPU cycles >>>on both 32-bit and 64-bit system. >> >>This is an interesting statement. What if the CPU has 4 megs of on-die cache >>and the hash table is 2 megs? >>Will it make a difference if the ram is 500 MHz DDR? Rambus? >>Does the CPU always wait that long period? >> >>Been a long time since I delved into timings, though I do know a memory fetch >>and a missed branch prediction are heavy blows.
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.