IP: from Bell RE: Bell v. Hillis: What about "Toy Story"?

[Dave Farber <farber () central cis upenn edu>]

From: Gordon Bell <gbell () microsoft com>
To: Dave Farber <farber () central cis upenn edu>,
        "pcw () access digex net (Peter Wayner)" <pcw () access digex net (Peter
Wayner)>,
        Jim Gray <gray () msg microsoft com>


The bet is far from being decided.  It takes some work to figure it out.   


Over a decade ago, Apollo made a film "Long Ray's Journey into Light" using 
all their workstations to do rendering!


The bet recorded in Hennessey Patterson is somewhat ambiguous <100 vs >1000. 
 The top central machines that Dongarra chronicles have over a Tflops of 
capacity and are all MPP-like e.g. Intel, CM5, IBM, T3D.  So, if these are 
generating lots of numbers I could lose unless I invoke the >1000 streams 
clause. Just to count the flops/month is hard, cause with MPP you often have 
to do lots of redundant calc in a node because it takes so long to get the 
numbers from a friend.  To  get MPPS to have high flops you often have to go 
to structures that, if you had finer grain machines, would be done at much 
lower flops rates.  Finally, to get high MPP flops you end up solving bigger 
problems that give you little in insight, but that take longer because the 
granularity is so large.  Run time goes as the cube or greater of a problem 
dimension.


When Danny described the CM5, my first comment was:  You didn't take the 
heads off the workstation did you?


At NSF I urged the CS community to work on parallelism vis a vis their 
workstations and to up the network to use them in situ.  Nothing happened 
till the NoW project and Chuck Seitz split off  Myrinet. Now at least things 
seem to be going the right way for the smart folks.


g
----------