[Beowulf] Berkeley View: A future in Parallel Programming?

[Mark Hahn]

>> Multicore: Today's dual and quad cores.
>> Manycore: many, many more cores per die [like Intel's recent prototype]

current multicore is just chip designers being lazy: everyone knows 
how to do traditional SMP machines, so hey, let's just put a system 
on a chip and call it innovation.  at least it uses up the embarassment 
of transistors bestowed by Moores's law.

another thread of the argument is from people who got burned by Intel's
clock-scaling jag.

both these problems seem like transient oddities where 
recovery is not any kind of major reorientation.

also, I have the impression that "petascale" is a major neurosis among 
the funding-gensia these days.  curiously, the View From Berkeley (VFB)
didn't contain this word...

>> If that is true, where does the Cell BE fit? Multi or Many?
>
> The Cell BE isn't an SMP chip, and neither is Intel's prototype. So
> I'd class it as Many.

afaikt, Intel's prototype doesn't really have a Grand Vision behind it - 
it's not a statement from Intel against coherent memory, just that the chip 
was an experiment in how many specialized FPUs you could put onto a chip
without relying on a lot of long wires.

> You don't have to change much to take advantage of Multi. Many is
> very different.

I admit VFB peeved me some in its "everything you know is wrong" tenor,
not to mention it's buzzword-competeness.