COMP:killer SuperDSP clusters, which ones?

Eugene Leitl (
Wed, 28 May 1997 10:41:06 +0200 (MET DST)

On Sat, 24 May 1997, r q watt wrote:

I thoroughly enjoyed these urls, this is honestly Good News. I wonder
however, whether a different brand of architecture might not offer a
better price-performance ratio than hypercubed 100 via MBit tx, PVM/PMI'd
Linux boxes.

The most natural topology for MD and lattice inspired runs is a
lattice/grid. The development has swung full circle: after the demise of
the venerable British transputer we now have diverse SuperDSPs with links,
noticeable numerical performance and lots of on-die core.

Consider the Analog Devices SHARC: a $100 chip in large quantities. It is
a 40 MIPS/120 MFLOPS, 1-4 MBIT on-die dual-ported SRAM 6 high-speed links
DSP. COAST SRAM modules have dropped down in price sufficiently that a
node can easily feature few MByte SRAM with 12-15 ns latency. TI's new
SuperDSP offers 1.2-1.6 GOPS performance, and 1 MBit core on-die, though
the first family member does not feature any high-speed links.

A 64-node superDSP cluster, assuming $200/node would fit in a mini tower
cabinet, and offer a _dedicated_ performance rivalling that of many
supercomputers for just 13 k$; make it 15 k$ for a Linux box to do I/O and
visualization. In case of a SHARC 64-node cluster we'd get 7 GFLOPS, a TI
would peak at 70 BOPS (!), making the latter's lack of an FPU negligeable
(one can do plenty of scientific things with scaled integers). This
numerical performance is a lot for 15 k$ -- and it's available for a
single user, for months if this needs be.

This stuff won't run UNIX, but there are now several DSP OSses around,
and message-passing code should be easy enough to isolate to make porting
to other platforms easy. You see, I am obviously very excited about this.

Now a question: is anyone already doing modeling on such beasties? If no,
is anyone going to? Is anyone aware of cheap multi-DSP cards for ISA/PCI
bus, coming with documentation and a compiler? I will search on the web,
however I'd like to first sample experiences of people who already worked
with above boards.

Please answer, I will summarize to the lists.


> homebrew Linux maspar supercomputers... or something that might
> evolve into that...
> and some fun software packages to run on them
> # try searching for molecular nanotech
> [ Will Ware's statement ]