Path: utzoo!attcan!uunet!cs.utexas.edu!sun-barr!olivea!oliveb!amdahl!JUTS!rbw00
From: rbw00@ccc.amdahl.com (  213  Richard Wilmot)
Newsgroups: comp.arch
Subject: Re: EFLOP architectures: when and for how much?
Summary: pipeline it
Keywords: pipelining
Message-ID: <1dbs02b=035v01@JUTS.ccc.amdahl.com>
Date: 29 Oct 90 01:40:33 GMT
Reply-To: rbw00@JUTS.ccc.amdahl.com (  213  Richard Wilmot)
Followup-To: comp.arch
Organization: Amdahl Corporation, Sunnyvale CA
Lines: 36

gcwilliams@watdragon.waterloo.edu (Graeme Williams) said in article
<1990Oct26.191032.9099@watdragon.waterloo.edu> :

> Namely :
> An instruction cannot be executed in a time shorter than the time
> it takes for a light beam to traverse the processing device.

> Thus if a single processor were to be capable of executing 10~18
> instructions per second - the physical size of the processor would
> have to be smaller than (3*10~8)/10~18 metres i.e. 0.3 nanometres.

> Unfortunately 0.3 nanometres is the same order of size as a single
> molecule - so either the processor is *EXTREMELY SIMPLE* :-) ,or it's
> not made of molecules! In short 10~18 operations/sec will probably
> be forever impossible for a single processor, indeed 10~15 probably
> is too.

While it is true that an instruction cannot be EXECUTED in less time
than light can traverse the computing device, nFLOPS are measures of
THROUGHPUT. We can improve throughput by some amount via pipelining of
the execution so that each section of a computing device performs only
part of the work for the instruction. After an initial delay to fill the
pipeline we could produce a result every 1/mth second where this time
between results < time for light to traverse the whole CPU.

Some people call this parallel computation and it is a form of parallelism.
Nearly all processors produced today are pipelined. Vector units are
usually (always ?) implemented with pipelining but they could include
true multiple execution units (e.g. adding 4 vector elements to 4 others
at the same time) to further increase throughtput.

I am usually more interested in bandwidth/throughput than the latency
of any individual instruction in a stream.
-- 
  Dick Wilmot  | I declaim that Amdahl might disclaim any of my claims.
                 (408) 746-6108