Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10.2 9/12/84; site desint.UUCP
Path: utzoo!watmath!clyde!burl!ulysses!mhuxr!ihnp4!houxm!whuxl!whuxlm!akgua!sdcsvax!sdcrdcf!trwrb!desint!geoff
From: geoff@desint.UUCP (Geoff Kuenning)
Newsgroups: net.micro.16k
Subject: Re: Corrigenda (24-bit addresses)
Message-ID: <351@desint.UUCP>
Date: Fri, 8-Mar-85 18:05:50 EST
Article-I.D.: desint.351
Posted: Fri Mar  8 18:05:50 1985
Date-Received: Mon, 11-Mar-85 11:26:51 EST
References: <794@sjuvax.UUCP> <5025@utzoo.UUCP> <2342@nsc.UUCP>
Organization: his home computer, Manhattan Beach, CA
Lines: 50

In article <400@terak.UUCP> doug@terak.UUCP (Doug Pardee) writes:

>> (a) 32 bits is a bit less than five Eagles worth of byte-addressed data.  I
>>     can easily concoct a microprocessor application where direct addressing
>>     of that data space is desirable, if not necessary.
>
>A *microprocessor* application????   Come on now, let's get serious.
>I _might_ grant that for an AI application you would need that kind
>of high-speed random-access memory, but no sane person would *ever*
>choose a 68xxx/32xxx type processor for AI.  You need a *lot* more
>MIPS than that.
>
>Let's see, using 256K DRAMs at $15 each, the cost of the RAM chips
>alone...

[there follows a complicated analysis of costs that concludes one would have
to expend $20 billion on RAM chips to get 4 Gbytes]

>So you're gonna spend maybe 20 G's or more on a memory system, and
>to "cut costs" you're gonna use an off-the-shelf microprocessor like
>a 68020 or 32032?  Gimme a break!

Give me a break yourself.  The part you quoted *explicitly* mentions the
Fujitsu Eagle, which is a disk drive.  Like I said before, I can easily
concoct a MICROPROCESSOR application that needs to address 32 bits worth
of MASS-STORAGE-RESIDENT data.  Surprise, surprise, it isn't going to be a
CPU-bound application.  It's going to be some sort of database-intensive
application.

It is true that there are other ways to handle disk-resident databases (or
other media;  trillion-bit memories on obscure slow media date back to the
late 60's) than as virtual memories.  Having done so, I can testify that
you will *always* wind up writing code that, in effect, simulates the
virtual memory hardware in software.  Sometimes you are lucky and you can
hide it pretty well.  But sometimes you have truly random reference
patterns that really do exhibit locality in short time spans, and then you
have to do it in software.  I know;  I've done it in software myself on
more than one occasion.  No thanks.

I might add that, even in a system that does not have virtual memory, a
wide address/data path makes writing software-virtual code much easier.
National got this part right, at least.  The funny thing is that you hear
pretty much the same noises from Intel as from National -- just replace all
occurrences of "24" by "16".
-- 

	Geoff Kuenning
	Unix Consultant
	(213) 545-4413
	...!ihnp4!trwrb!desint!geoff