[rescue] SMP on intel wasteful?

[Joshua D Boyd]

On Wed, Jun 26, 2002 at 09:10:50PM -0400, Joshua D Boyd wrote:
> > >  Umm, WHAT?  Show me one vector machine that ISN'T a "specialized SIMD
> > >approach".
> > 
> > Aren't "Vector Processor" and "SIMD processor" synonymous?
> 
> No.

That was an awefully short answer.  Let me provide a more friendly one.

A SIMD processor has numerous processing units that execute the same
instruction on different data (hence Single Instruction Multiple
Date).

Vector processors are considered a subclass of this.  While the data
in the SIMD pipeline can be different (meaning the meaning of the
data), in vector processors, the data is expected to have the same
meaning.

Basically it comes down to vector processors want to do strict vector
work (in the linear algebra sense) while SIMD machines are happy to do
that, but they are also happy to work on data that doesn't strictly
meet the definition of a vector.  While vector machines can do this,
all the special cases needed will really slow things down a lot.

So, vector machines are great for solving PDEs, but are't so hot for
many audio, video, and 3D tasks.  A general SIMD machine would
probably be faster at audio and video processing, but it isn't as
optimized for strict vector operations.

Also, I get the impression that on vector machines, the size of the
vectors can be quite large, but most simd architectures I've seen seem
to limit the "vector" size it will process to 8 elements.

If this sounds kinda hazy, it is mainly because I don't have handy my
definition of what a strict vector is.  I know that [x y z w] doesn't
count, mainly because of the w element.  And I know that packing two
sound samples together like [right-chanel1 left-chanel1 right-chanel2
left-chanel2] doesn't count because the right will usually mean
something different.  [right-chanel1 right-chanel2 right-chanel3
right-chanel4] would count though.

-- 
Joshua D. Boyd