numbers

["David J. Littleboy" <Littleboy@YALE.ARPA>]

A reply to Mike Fischer's comments:

A brhef clarification:  At the  current  time  I  do  not  see  Lisp  machines
replacing  Apollos.   A  department  may  be  able  to afford $10K to $15K per
student, but it can't afford $50,000.  I see Lisp machines as an  alternative
for people whose programs are too large to run on Apollos.

       1.  What is the cost per node for all that speed?  If a node that
       costs twice as much runs 3-5 times as fast, is that really a good
       deal, given that most of the time it sits practically idle doing
       text editing?

Some  of  us  need more cpu cycles than others.  In 1974, ten years ago, I had
access to a KL-10 (read "2060") evenings with few enough  users  that  it  was
efectively  my  own.  Two years ago, one could come in in the evenings and run
on a similarly empty research machine.  Today my T programs run 10 times  more
slowly when they are not garbage collecting, and many orders of magnitude when
they are.  For compute intensive research (e.g. machine learning  experiments)
the  apollos are a joke.  The tern, at 3 times faster, will continue to be one
third the cycles available ten years ago.  So much for progress.

In terms of cost, IFF Symbolics is willing to give us the kind of deals I was
quoted at AAAI this summer, 3640s should be cost competitive on a node per node
basis with terns. My estimate is that a 3640 OUGHT to be available at
under $50,000 per node.

       2.  What is the true cost of switching systems, once you factor in
       the additional staff needed to support it and the increased difficulty
       of interaction with others in the local community?

The point of the 3600's would be that, being for people who need cycles, those
people  would  be  responsible for figuring them out for themselves.  However,
since Drew maintains a NISP and a DUCK for the 3600, all his students can  run
their  code  without  change.   Since 3600's have a company behind them, not a
graduate department in which each group would love to see the others go  away,
software support should not be a problem.  Hardware support has been quoted at
$7,000 per node-year:  admittedly expensive.

        3.  How good is the software?  Does it support transparent remote file
        access?  Mail?  TeX?  Scribe?  Graphics?  Yale-style editors?  C?

It  has  a  lot  of software.  It is claimed to support mail, transparent file
access, Zmacs (an Emacs style editor) (come on now, you aren't going to refuse
to let someone else use a system just because it doesn't support your favorite
editor, are you??), FORTRAN, and graphics, and  might  support  scribe  (don't
hold your breath).

Although  it  takes  getting  used  to,  the  environment  is quite nice.  The
compiler is fast enough that exiting  from  the  editor  can  go  through  the
compiler  and  still  be  reasonably fast:  at about the cost of swapping in a
different "context" on a diskless node, one can debug compiled code.

           4.   How  good  is  the  Lisp?  Does it have the clean semantics to
           facilitate the development of large systems that T has?   How  long
           does  it  take  to  learn  to  use  effectively  (i.e.  what is its
           "kludginess coefficient")?  How many of our  faculty  and  students
           have the time to invest in learning to use it effectively?

Zetalisp  is not T.  Not by a long shot.  It feels like returning to the stone
age.  HOWEVER, it  is  possible  to  work  in  it,  and  it  does  support  a
methodology, admitedly on the kludgey side, for developing large systems.


         5.  What is its growth path, i.e. if you take similar  benchmarks  in
         two years' time, what will you expect to see?  Is there any reason to
         believe the Symbolics class of machines will remain faster  than  the
         Apollo class?  Maybe I hold out too much hope for the new T compiler,
         but I had the impression that it might yield a speed increase of  two
         to three times.  If so, the Tern running T would already approach the
         3670.

First of all, neither terns NOR the new T compiler exist.  Apollo has  stopped
accepting  orders for terns, and although the future promises DN300 size boxes
which are 6Mb Terns, that depends on both 256k chips and the 68020, neither of
which  are  available  yet.  The 3600 is here.  Now. It is also a TWO YEAR OLD
machine; we saw a functioning prototype at AAAI two years ago.  T 3.0  was  to
have  been  here  after  a  full  summer  of  work by all involved.  Now it is
promised for December. But most of the relevant people are no longer working on
it. (assigning blame here is ugly, so don't misinterpret me: I know its an
unfortunate situation)

Both Symbolics and Apollo have  new  machines  in  the  works.   There  is  an
instruction  pre-fetch  unit  which  may speed up 3600s 20% to 50%.  There are
semi-custom VLSI things in the works also.  But even if the Tern can catch  up
in  speed  with  the  3600,  the 3600 has a 27 bit cons cell address space, as
compared to 19 bits in the 68010, and 22 AT BEST IFF they pin out the existant
address  lines  in the 68020.  (am I wrong here?  Could the 68000 architecture
support 32 bits of byte address, and thus 27 or 28 bits of cons cells???)

--Mike

-------
-------