🚀 go-pugleaf

In article <1992Dec10.164402.17832@news.eng.convex.com>, dodson@convex.COM (Dave Dodson) writes:
|> In article <1992Dec10.141843.2836@cis.uab.edu> hyatt@cis.uab.edu (Robert Hyatt) writes:
|> >I suspect that many forget that there is a second important part to the
|> >precision question.  While 32 bits might work for some applications, *if*
|> >the application needs a supercomputer, it is probably going to iterate 
|> >enough to wipe out most of the 32 bits of accuracy they need.
|> 
|> The oil industry consumes lots of supercomputer cycles doing seismic
|> analysis, and 32 bits is more than enough precision for that.  After
|> all, the seismic data has only 2 or 3 significant digits.  Using more
|> than 32 bits doesn't improve their ability to find oil. 

Much of the routine seismic processing in the oil industry indeed
does not require more than 32 bits (and not even really that 
many - much of the data is originally recorded using only 24 bits.)
Many of the algorithms used in processing the data are not iterative 
at all, but simply apply linear operators (often multidimensional) 
to the data.  The data sets used can be very large (from 10^8 to 10^11 
samples or so per survey and getting even bigger).  I/O bottlenecks 
moving that much data around are often much more of a problem than 
loss of precision.  

For seismic use, it is desirable for machines using 64 bit 
words to provide very efficient support for packing two 32 bit 
floats per word (both in memory and on disk), as well as the ability 
to do arithmetic using these packed floats without incurring major 
overhead cost or programming headaches.  Without such packing, using 
64 bit words can potentially mean doubling the memory and disk 
requirements, the cost of which can easily already dominate the 
price of a computer for seismic processing.

Seismic processing is not the only use for supercomputers in the
oil industry, but it is a major one, and constitutes a large enough
market segment that supercomputer manufacturers are generally aware 
of its particular needs.  I do hope that people who are designing
64 bit microprocessors, and particularly those who want to build 
MPP machines with them, remember that, for some of us, 32 bit 
floats are quite adequate much of the time.  

------------------------------------------------------------------
 Paul Fowler                     Western Geophysical R+D
 fowler@de01.denver.waii.com     Englewood, Colorado