[rescue] RE: divide and conquer? (CPU design)

[Ken Hansen]

I am *far* from being an EE, but I do know that once you move componentls off the same die, peformance takes a *huge* hit - witness the PPro with it's on-die cache compared with similar chips with cache on a separate die (PII?).

There are somethings in life where I defer to the engineering department, like high-speed computing and automatic transmissions*. I look to Cray, and see that their latest designs started to use (IIRC, Dave McGuire feel free to correct me) collections of "off-the-shelf" CPUs as opposed to dividing the CPU into numerous components, as you describe.

Real gains in speed are achieved by proximity, and dividing a CPU into multiple pieces of silicon (or whatever) will have a negative effect.

That's my $.02,

Ken


-----Original Message-----
<snip>

Message: 5
From: dave at cca.org
Date: Wed, 20 Jun 2001 23:38:03 EDT
Organization: The Center for Computational Aesthetics
To: rescue at sunhelp.org
Subject: Re: [rescue] 670MP windfall in the surplus diving frenzy today.....
Reply-To: rescue at sunhelp.org

mcguire at neurotica.com writes:

<snip>

Any application where there is a major multi-tasking load, and 
system throughput is more important than single-process performance,
a single chip of [n] million transistors would be much more usefull 
if it were a bunch of little simple CPUs than one big complex one.

Put another way: given, say, eight instruction units on a chip,
you will *not* be able to keep them occupied trying to dynamicaly
schedule a single instruction stream. The OS on a heavily-loaded
server *will* be able to keep them fully occupied if they appear
to be eight seperate CPUs.

If you want to get really insane, you share resources between
the CPUs, which gets you back to the Barrell Processor design.
(The only example of that that I can remember offhand was the IO
processors in the CDC-6600. Anyone know of any others?) But then
you're getting excessively complex again.

Cache design becomes significant (both how best to split up the
real estate, and how to deal with coherency), but that's solvable.

And of course, you can take [n] of these chips and make a really
big SMP system...

I'm sure other people have thought of this, and are probably already
working on it.

-- david fischer -- dave at cca.org -- www.cca.org -- Cthulhu told me to. --
<snip>