[rescue] Cray J90s

[Dave McGuire]

On June 12, dave at cca.org wrote:
> >  Vector processors, on the other hand, operate on vectors as atomic
> >values.  They can perform the same operation on whole lists of numbers
> >(vectors) in one operation...one vector instruction.
> 
> Which is not to say one clock cycle or anything close. In fact,
> isn't there usually just one or two instruction units that are
> processing each element of that array? So - it saves you the 
> overhead of keeping track of the loop, and optimizes utilization
> of the pipeline. It doesn't actually calculate all the entries
> simultaneously.

  Also known as: extremely deep pipelining.  While this is the case
for most Cray designs, it is not necessarily the case for all vector
supercomputers.  Whether or not the individual vector element
operations are accomplished truly simultaneously is a matter of
implementation, not a matter of being a vector machine or not.

> And for a bizarre comparison: couldn't the Cyber 205  and its ETA
> descendants do memory-to-memory vector instructions on 64K
> element arrays?

  They have no vector registers...they are all memory-to-memory vector
machines.  Unless you're specifically referring to the vector length,
which I don't recall offhand.

          -Dave McGuire