Intel compilers and libraries

Robert G. Brown rgb at phy.duke.edu
Thu Oct 9 09:16:43 EDT 2003


On Thu, 9 Oct 2003, C J Kenneth Tan -- Heuchera Technologies wrote:

> > 4) Put your performance whitepapers on your website, or it looks
> > fishy.
> 
> Our white papers are not on the Web they contain performance data, and
> particularly, performance data comparing against our competitors.  It
> may expose us to libel legal issues.  Putting legitimacy of any legal

Expose you to libel suits? Say what?

Only if you lie about your competitor's numbers (or "cook" them so that
they aren't an accurate reflection of their capabilities, as is often
done in the industry) does it expose you to libel charges or more likely
to the ridicule of the potential consumers (who tend to be quite
knowledgeable, like Greg).

One essential element to win those crafty consumers over is to compare
apples to apples, not apples to apples that have been picked green,
bruised, left on the ground for a while in the company of some handy
worms, and then picked up so you can say "look how big and shiny and red
and worm-free our apple is and how green and tiny and worm-ridden our
competitor's apple is".  A wise consumer is going to eschew BOTH of your
"display apples" (as your competitor will often have an equally shiny
and red apple to parade about and curiously bruised and sour apples from
YOUR orchard) and instead insist on wandering into the various orchards
to pick REAL apples from your trees for their OWN comparison.

What exactly prevents you from putting your own raw numbers up, without
any listing of your competitor's numbers?  You can claim anything you
like for your own product and it isn't libel.  False advertising,
possibly, but not libel.  Or put the numbers up with your competitor's
numbers up "anonymized" as A, B, C.  And nobody will sue you for beating
ATLAS/GCC/GSL numbers -- ATLAS etc are open source tools and nobody
"owns" them to sue you or cares in the slightest if you beat them. The
most that might happen is that if you manipulate(d) ATLAS numbers so
they aren't what real humans get on real systems, people might laugh at
you or more likely just ignore you thereafter.

What makes you any LESS liable to libel if you distribute the white
papers to (potential) customers individually?  Libel is against the law
no matter how, and to who, you distribute libelous material; it is
against the law even if shrouded in NDA. It is against the law if you
whisper it in your somebody's ears -- it is just harder to prove.
Benchmark comparisons, by the way, are such a common marketing tool (and
so easily bent to your own needs) that I honestly think that there is a
tacit agreement among vendors not to challenge competitors' claims in
court unless they are openly egregious, only to put up their own
competing claims.  After all, no sane company WOULD actually lie, right
-- they would have a testbed system on which they could run the
comparisons listed right there in court and everybody knows it.  Whether
the parameters, the compiler, the system architecture, the tests run
etc. were carefully selected so your product wins is moot -- if it ain't
a lie it ain't libel, and it is caveat emptor for the rest (and the rest
is near universal practice -- show your best side, compare to their
worst).

> issues aside, it is not good for any business to be engulf in legal
> squabbles.  We are in the process of clearing this with our legal
> department at the moment.
> 
> As I have noted in my previous e-mail, anyone who wants to get a hold
> of the white papers are welcome to please send me an e-mail.

As if your distributing them on a person by person basis is somehow less
libelous?  Or so that you can ask me to sign an NDA so that your
competitors never learn that you are libelling them?  I rather think
that an NDA that was written to protect illegal activity be it libel or
drug dealing or IP theft would not stand up in court.  Finally, product
comparisons via publically available benchmarks of products that are
openly for sale don't sound like trade secrets to me as I could easily
duplicate the results at home (or not) and freely publish them.

Your company's apparent desire to conceal this comes across remarkably
poorly to the consumer.  It has the feel of "Hey, buddy, wanna buy a
watch?  Come right down this alley so I can show you my watches where
none of the bulls can see" compared to an open storefront with your
watches on display to anyone, consumer or competitor.  This is simply my
own viewpoint, of course.  I've simply never heard of a company
shrinking away from making the statement "we are better than our
competitors and here's why" as early and often as they possibly could.
AMD routinely claims to be faster than Intel and vice versa, each has
numbers that "prove" it -- for certain tests that just happen to be the
tests that they tout in their claims, which they can easily back up.
For all the rest of us humans, our mileage may vary and we know it, and
so we mistrust BOTH claims and test the performance of our OWN programs
on both platforms to see who wins.  

I'm certain that the same will prove true for your own product.  I don't
care about your benchmarks except as a hook to "interest" me.  Perhaps
they will convince me to get you to loan me access to your libraries etc
to link them into my own code to see if MY code speeds up relative to
the way I have it linked now, or relative to linking with a variety of
libraries and compilers.  Then I can do a real price/performance
comparison and decide if I'm better off buying your product (and buying
fewer nodes) or using an open source solution that is free (and buying
more nodes).  Which depends on the scaling properties of MY application,
costs, and so forth, and cannot be predicted on the basis of ANY paper
benchmark.

Finally, don't assume that this audience is naive about benchmarking or
algorithms, or at all gullible about performance numbers and vendor
claims.  A lot of people on the list (such as Greg) almost certainly
have far more experience with benchmarks than your development staff;
some are likely involved in WRITING benchmarks.  If you want to be taken
seriously, put up a full suite of benchmarks, by all means, and also
carefully indicate how those benchmarks were run as people will be
interested in duplicating them and irritated if they are unable to.

   rgb

-- 
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



_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list