NetNews Offline 2/NetNews Offline Volume 2.iso/news/comp/sys/amiga/programmer/1890

Internet Message Format | 1996-08-05 | 3.4 KB

Path: news.delcoelect.com!c23jwd
From: [e-mail] (Jeffrey William Davis)
Newsgroups: comp.sys.amiga,comp.sys.amiga.hardware,comp.sys.amiga.programmer,no.amiga
Subject: Re: FWD: Fate of 68080
Date: 24 Jan 1996 20:24:00 GMT
Organization: Delco Electronics Corp.
Distribution: usa
Message-ID: <4e64h0$[e-mail]>
References: <4d3c27$[e-mail]> <[e-mail]>
NNTP-Posting-Host: koptsw19.delcoelect.com

In article <[e-mail]>,
Kenneth C. Nilsen <[e-mail]> wrote:
>>The current/near future top-of-the-line is the PPC 604@150 MHz.
>>The 300 MHz was foreseen for the 64-bit PPC 620, but the situation about this
>>beast is unclear since the projected performance gain is not enough these
>>days. Instead a design rework will be done, maybe they call it 630 or so.
>>Anyway the race for MHz sounds silly to me, since the memory can't keep pace
>>with it, except for large (and expensive) caches. 700 MHz is science fiction
>>I'd say.
>
>Yeah, you got a point. But wouldn't it be possible to write data parallell to
>memory, I mean, using double memory bus writes so that in principle you can
>write not 1 byte, but 2 bytes simultainously or even 4 bytes etc. (hope you
>get the picture) ?

This 'principle' which you loosely describe is already commonly done, and
64, 128bits (16 bytes) or more are written and read simultaneously.
Writing data has never really been the problem, it's READING it that
causes a bottleneck that is tough to get around.

When writing, the data can be handed to a cache that will write it whenever
it has the time.  If a read is attempted on that data while it is still
in the cache, you simply deliver (read) the cache value.  Granted, there
are numerous ways to get data into the cache (including reads) but that's
not important here.

When reading, the processor expects the data NOW.  There is a limited
amount of time between when the processor is able to give an address and
when it expects the data to be ready.  If you take longer than this to
retrieve the data, the processor has to wait; hence, wait-states.

With RAM technology rapidly falling behind processor speed, we need to
somehow 'predict' which piece of data the processor will want and (at
the very least) begin reading it before the processor even asks for it.
Then place it in a device (cache) which can deliver it to the processor
more quickly.  This is where the 'expensive' cache comes in.

CPUs having asynchronous address and data busses make it easier to
manage the high MHz, since it takes less time to latch an address than
it does to complete a R/W cycle on the data bus.  This allows you to
grab multiple addresses and begin working on them simultaneously while
the data busses are saturated; continually trying to keep the data bus
from ever having to wait.

Unfortunately, the only perfect prediction would be one that could
deliver any piece of memory on demand as fast as the processor can
accept it - hence the < 1ns DRAM @ 700MHz!  Not likely to happen in
the near future.  The further the CPU MHz fly past the current RAM
technology, the more rediculous it becomes.

--
=======================================================================
Jeffrey W. Davis [phone]   Domain: [e-mail]
Software Engineer                UUCP:   deaes!c23jwd
Delco Electronics Corporation    GM:     8-322-0503         Mail: CT40A