[prev in list] [next in list] [prev in thread] [next in thread] 

List:       beowulf
Subject:    a DSP-based supercomputer
From:       Rahul Dave rdave () central ! cis ! upenn ! edu
Date:       1998-04-23 15:45:28
[Download RAW message or body]

> 
> However, I think that the price tag would, as I said above, be
> CONSIDERABLY higher, as much as a factor of ten higher.  On the third
> hand, the system could be used for a lot more than QCD calculations and
> indeed subsets of the system could be easily used for different
> calculations at the same time.  The Columbia system is really not very
> good for general purpose calculations (too little memory per node and a
> predefined toplogy, although one can rearrange the topology without
> having to redesign the whole thing) and isn't necessarily all that great
> even for heavily IPC bound parallel calculations.  It is best for
> problems that match the (possible) IPC topology(s), e.g. lattice
> simulations.
> 

A programmable routing chip could act as a co-processor to the CPU chip, and
dynamically change topology based on the problem..say , where your soliton
moved to, for example, or where a vortex was formed in a hydrodynamic code.
The chip would be programmed via a header in your communication packet.

We have a project here at UPENN to program ethernets by embedding code in
the IP headers. We are going to try this stuff out on the Beowulf we're
building here. The idea can be scaled down to the hardware level I think.

> > 
> > It's certainly... ambitious... to be able to run so many processors
> > instead of fewer, faster ones.
> > 
> > -- g
> > 

Its interesting to read the last chapters of Hillis's thesis, the Connection
Machine, where he cries from the belltowers about building computers that
mimic models of physical computation.

Now, who is selling me one of them PCI cards?

Rahul

[prev in list] [next in list] [prev in thread] [next in thread]