Hardware Progress: $397 (fwd)

Robert G. Brown rgb at phy.duke.edu
Wed Mar 27 16:12:09 PST 2002


On Wed, 27 Mar 2002, Bob Drzyzgula wrote:

> A couple of years ago I stared at Intel's historical price
> and performance numbers numbers until I started to hallucinate
> and believe that I saw some patterns there. I did a few
> calculations, some of which many of y'all might find
> interesting. I put it up at
> 
> http://www.drzyzgula.org/aicpd.pdf

Sheer genius.  Is this, by any miracle, in latex?  Can I, by any chance,
add this as an appendix or chapter or whatever to my online book?

Lovely job.  Of course yes it would be great to have references and so
forth as you mention below (not to mention some GRAPHS, since some folks
have a hard time reading tables of numbers and since quality of fit is
easier to judge on a graph:-).  Still, superb work and bound to be
helpful to anybody who ever needs to do a SERIOUS CBA for total
integrated capacity for a compute cluster.

I stand in awe.

    rgb

> 
> A couple of notes:
> 
>  * I pretty much just use an arbitrary 18-month Moore's
>    Law doubling time without much comment. However, I had
>    sufficient data to float this doubling rate and have the
>    least squares pick it for each pricing level from $150
>    to $800. When I did this, the doubling period was more
>    like, IIRC, 15-17 months. Interestingly, although not
>    surprisingly, the doubling period was shorter for less
>    expensive processors. As a result, the model showed
>    faster (in the SPECint sense) processors ultimately
>    being less expensive than slower processors. To some
>    extent, this makes some sense in that really expensive
>    x86 processors tend to have larger caches which drive
>    up the price without adding to CPU speed. This doubling
>    skew was a hassle for the rest of the paper, however,
>    so I fixed it a single value across all pricing levels.
>    I tried a 16 month period, but it didn't change the result,
>    so I went back to 18 months because the economists I
>    was trying to reach were more familiar with the 18 month
>    mythology.
> 
>  * I wrote this paper for internal use at my office.
>    Although it's not sensitive or anything, I probably
>    would have done a bit better job with references and
>    such if I had planned on publishing it. A large portion
>    of the data was collected by Micro Design Resources
>    and/or the SPEC consortium.  I supplemented this through
>    other numbers I pulled off the web from things like
>    the Chip Directory and Intel's press releases. This
>    all should have been credited.
> 
>  * Internally, I had an appendix that contained all the
>    numbers and the Octave program I used to do the
>    calculations. This appendix isn't available because much
>    of the data collection is copyrighted. However, if anyone
>    is really interested in playing with it -- or especially
>    updating it -- let me know and I'll see what I can do.
> 
>  * I'm not a statistician. This isn't my area of expertise,
>    so there could be some egregious errors. I had some
>    real statisticians read it, though, and they didn't
>    barf or anything. Still, if anyone spots anything
>    suspicious feel free to let me know.
> 
> --Bob
> 
> On Wed, Mar 27, 2002 at 05:20:54PM -0500, Robert G. Brown wrote:
> > 
> > > I would think a cost reduction by 2-5x maximum
> > > in 2006.
> > 
> > More practically, microprocessor manufacturers and other tech component
> > manufacturers don't generally focus on reducing consumer costs, they
> > focus on increasing speed (and other overall performance measures) at
> > constant consumer cost and relatively high margins.  They can do this
> > because tech profit margins depend to a significant extent on demand in
> > a fairly inelastic market with very limited competition.
> > 
> 

-- 
Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu






More information about the Beowulf mailing list