Path: utzoo!utgpu!watmath!att!tut.cis.ohio-state.edu!cs.utexas.edu!uunet!crdgw1!sungod!davidsen
From: davidsen@sungod.crd.ge.com (ody)
Newsgroups: comp.sys.ibm.pc
Subject: Re: 80386 memory expansion board needed
Keywords: 16 bit memory on 386 bus question
Message-ID: <1540@crdgw1.crd.ge.com>
Date: 8 Aug 89 18:28:40 GMT
References: <35950@bu-cs.BU.EDU> <1151@mitisft.Convergent.COM>
Sender: news@crdgw1.crd.ge.com
Reply-To: davidsen@crdos1.UUCP (bill davidsen)
Organization: General Electric Corp. R&D, Schenectady, NY
Lines: 37

In article <1151@mitisft.Convergent.COM> burton@mitisft.convergent.COM writes:

| A somewhat related question is whether or not you can take 16 bit memory boards,
| say an AST ADVantage, with 120/150 ns memory, and run it successfully on a
| 20 or 25 MHz system.  Is this possible?  How much does performance suffer?
| Does a RAM cache on the motherboard help any?

  I've been running a mixed 32/16 bit system for several years and have
done quite a bit of testing. Here's how it works. If you're running DOS
and using the high memory for RAMdisk, the cache won't help much, if at
all. If you run QEMM or other software to use the extended mem as
expanded mem, it will help quite a bit.

  To get the average access penalty you can perform a calculation like
this: access to fast memory = 2 cycles for 32 bits, slow memory 2 cycles
+ 1 wait state (16MHz, 2 for 20, 3 for 25, using 120ns memory). Since a
fetch takes two accesses on the 16 bit bus, double that. I used 85%
cache hits, because (a) I saw that number in a lot of places, and (b) I
measured 84.4% on a simulator using a C compile and an awk program.

		<_ cycles ->	< effective >
CPU MHz		fast	slow	clocks	waits
16		2	 6	2.60	0.60
20		2	 8	2.90	0.90
25		2	10	3.20	1.20

  From this I conclude that a 20MHz 386 running 64kb cache and memory on
a 16 bit bus will be about as fast as an 1w/s machine without cache. The
penalty will be less for 16 bit data transfers (all instructions are 32
bits, I believe).

  These figures are for AMI 16 MHz 386, running Xenix/386. Any other
hardware or software will change them (but not much). These calculations
reasonably match actual tests I have made.
	bill davidsen		(davidsen@crdos1.crd.GE.COM)
  {uunet | philabs}!crdgw1!crdos1!davidsen
"Stupidity, like virtue, is its own reward" -me