Andrew Ball wrote to James Bradley, "Parallel vs. Serial"

 AB> Hello James,

  JB> RE: Transport/HyperTransport
    > That's what it is called then? We aren't talking about
    > Symmetric Multi Processors, but with separate MB-CPUs
    > working on the same problem?

 AB> You've just mentioned three different things:-

That's why I thought I'd ask. 
 AB>   HyperTransport
 AB>     AMD's high-speed serial bus for connecting various
 AB>     things (potentially including multiple processors) that
 AB>     all live on the same circuit board.

 AB>   Symmetric Multi Processing (SMP)
 AB>     A (perhaps slightly crufty) approach to parallel
 AB>     computing: shove two or more microprocessors onto the
 AB>     same bus - they share memory, peripherals etc.

 AB>   "Separate MB CPUs"
 AB>     Sounds more like what seems to be called "clustering"
 AB>     these days: shove a rack full of inexpensive machines
 AB>     (which may or may not feature SMP) and link them
 AB>     together using something like a gigabit LAN.

Clustering was the word I was looking for. When you say "these
days" suggests
there is a more traditional term for it? 
 AB> The idea introduced to me as "waferscale" could, in theory
 AB> give you the same raw processing power as a rack full of
 AB> conventional PCs, on something the size of a cookie.  Doing
 AB> that would introduce some I/O challenges, and providing a
 AB> useful amount of memory for each processor core would be non
 AB> -trivial.  That might become more practical once we're able
 AB> to build circuits on (or rather in) something that isn't
 AB> limited to two dimensions, as a semiconductor wafer is.

Ya... Wouldn't that be nice? I trust the cube-type memory experiments haven't
proven too fruitful? (To clarify: I read some years ago about a polymer 
that would store data via three lasers focused onto one part of the matrix.)
  JB> For simplicity sake, would one 'puter work on part A
    > of the math, while the second works on part B?

 AB> Yes, but some problems are more tricky than others because
 AB> if "part B" depends on the result of "part A",
there's some
 AB> latancy involved.  If you have a steady stream of identical
 AB> calculations, that may not be much of a hit.

Same problems with the super-scalar PowerPC?

 
___ MultiMail/Linux v0.45

--- Maximus 3.01