Author: Robert Hyatt
Date: 09:17:09 06/25/02
Go up one level in this thread
On June 25, 2002 at 06:07:57, Vincent Lejeune wrote: >On June 24, 2002 at 22:29:38, Robert Hyatt wrote: > >>On June 24, 2002 at 13:31:20, Bo Persson wrote: >> >>>On June 24, 2002 at 10:30:26, Robert Henry Durrett wrote: >>> >>>>This addresses half the problem. What if the microprocessor wishes to WRITE >>>>something. Why not write it directly to a huge cache and bypass RAM entirely? >>> >>>It does, sort of. >>> >>>The processor writes to the cache, which will update the RAM *eventually*. The >>>processor doesn't have to wait for the update, so if you're lucky you will not >>>see the delay. >>> >>>Current processors even have write-buffers queueing data going to the cache... >> >>Most use "write back" or "copy back" which means that memory is not updated >>until the modified cache line gets replaced by something else. At that point, >>the modified (dirty) line is first flushed back to memory before it is replaced >>by something new. With luck, this turns a bunch of memory write operations into >>a bunch of cache writes with one memory write later on... >> >>And then there is "victim cache" to hold stuff that was "displaced" for a bit >>(from cache) in case it is needed again soon. >> >> >> >>> >>>>If you had extremely large caches, couldn't RAM be dispensed with entirely? >>> >>>Or, if you had fast enough RAM, caches could be dispensed of. :-) >> >> >>Micron used to do this. They used to have the fastest 386/486 boxes running. >>Their entire memory was SRAM rather than DRAM with an SRAM cache. Made them >>very fast. And much more expensive. Of course, 1 gig of SRAM would be very >>big compared to 1 gig of DRAM. > >The improvement of "all Sdram" souldn't be very big, if I remember correctly , >at this very old time (386/486) the figure was : 80% of the access memory were >in the cache with 16 KB and 95% for 256 KB cache ! >The typical amount of ram was between 4 MB and 16 MB > It depends. Some folks used those machines for number-crunching, even at those clock rates. With large arrays, cache was blown out instantly as the programs generally iterated over the entire array. The SRAM memory could keep up better, at a price ($). We had one math person here that bought one and liked the results, if not the price. > > > >> >> >> >> >>>That was actually the case for micros until about 15 years ago, when caches were >>>introduced alongside the "extremely" fast 486. >>> >>>A couple of years before that, 120 ns RAM matched an 8 MHz 286 processor pretty >>>well. >>> >>> >>> >>>> >>>>Bob D.
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.