[NCLUG] PC benchmark pointers?

[Matt Rosing]

>Hey, without getting into a discussion of the value of benchmarks,
>can anyone point me to a comparison of PCs to supercomputers?

One benchmark commonly used is linpack (a dense matrix solve). It's
also used as the basis for sorting the TOP500 (machines) list at UT
(Tennessee). If you look at the list, the top 6 are based on
workstations (IBM, SGI,...). The highest measured performance is
~5Tflops built from 8192 nodes. There are some other machines not
based on workstation chips like NEC, Fujitsu, and Hitachi. They use
vector chips for each node, I believe, but still have to group a lot
together.  There is a beowulf down at #84. 

I don't think there's a lot of difference between the basic chips of a
lot of these machines. The main difference is in the interconnects
between the nodes. So, to answer your question, if you tie enough PCs
together the right way, you can make a good supercomputer.

BTW, there's a machine being built in Japan that is expected to get
close to 5 Peta-flops. It's all custom, though.

I'm really hoping beowulf's take hold. I work on these machines all
the time and, considering all the money that's spent on them, they can
be worse than using card decks.

>I'm particularly interested in how desktop PCs these days compare to
>supercomputers of yore like the Cray XMP.

The XMP isn't even on the list of 500. I believe you can get a longer
list of linpack results. I've seen benchmarks with Dell 200Mhz chips. 
check out http://netlib2.cs.utk.edu/benchmark/top500.html. Also just
look at http://netlib2.cs.utk.edu/benchmark for a bigger list of
benchmarks. 

Matt