rsf%diablo@sri-unix.UUCP (01/26/84)
From: Ross Finlayson <rsf@diablo>
Gerry Sussman gave this talk at Stanford on January 24th. Here was the
abstract:
---------
A Digital Orrery
Gerald Jay Sussman
CalTech Theoretical Astrophysics Group
MIT Artificial Intelligence Laboratory
The Orrery is a computer specifically designed for doing high
precision orbit integrations at blazing speed. It is intended to be
used as a back-end processor, to be attached to a small conventional
host computer (eg. an IBM PC). The host computer will be used to set
up and access the states of the particles, and to set up the control
sequences for the Orrery.
The Orrery is made of a number of planet machines controlled by a
central controller. For N bodies there are N planet machines hooked
in a circle, such that data can be sent from machine i to machine
(i+1)mod N. This configuration allows the Orrery to perform
integration steps in O(N) time (with O(N) hardware). The machine has
a SIMD controller which broadcasts identical instructions to each
planet machine. There are no data dependent steps in the microcode,
so the SIMD controller needs no inputs from the planet machines.
I will discuss the problems the Orrery is being built to solve, the
current state of and the details of the design, and the plan for
construction.
-------
rsf - Additional points that I picked up from the talk: A prototype machine
is currently under construction; testing should begin sometime around June.
The machine is being built with "off the shelf" TTL (there's no custom
VLSI), including a special HP floating-point processor chip (I forget the
details of this). The completed machine will be ideally suited for the
solution of N-body problems, where 'N' is fairly small (say < 10). An example
would be the computation of the influence (over several thousand years)
of the Sun, Mars and Jupiter on the orbit of a particular asteroid. Sussman
pointed out that even such problems with small 'N' have no analytic solution
in general; furthermore, they cannot be easily 'vectorized' for efficient
solution on a machine such as a Cray. Sussman expects (typically) a
billion-to-one speedup over "real life". That is, it would conceivably be able
to simulate one billion years of a planet's orbit in roughly one year of
machine time.
The machine will not be suitable for solving problems for very large 'N'
(eg for globular clusters). Such problems could perhaps best be treated as
problems in fluid mechanics instead.
The machine will not be 'hardwired' for simple Newtonian mechanics (GMm/r**2).
With appropriate hacking, the central controller's microcode could be modified
so that (for example) tidal, drag and relativistic effects are also taken into
account.