🚀 go-pugleaf

Spice Benchmark results

guy fedorkow, Nov 25, 1985 at bbn

    In the process of qualifying a machine for a vlsi
workstation, spice benchmarks were run to compare Sun, DEC and Masscomp
machines.  Three spice decks were used, with run times from under
one to about 5 minutes.  The two circuit configurations used
contained 17 and 19 transistors.  In the third test, the default
gate length parameter was made larger to cause the simulation
to fail to converge.


    Tests were run on a VAX 11/785, Sun3/160C, the MicroVax II and
two Masscomp models, the 5400 and 5700.  Both the sun and the masscomp
offer 68020/68881 processor/floating point combinations.  The
masscomp machine claims to be running the '881 at 16 MHz, while the Sun
claims only 12 MHz for the '881.  Both run the '020 at 16 MHz.

Tests on the MicroVax II were run at UCB by George Jacobs.

    The tests on the VAX/785 were run in the early morning at BBN,
when the load factor was under 0.30.  Time values cited are total
elapsed time, as measured by the 'time' command.  All times are
expressed in seconds.

All the machines tested had at least 4 Mbytes of memory.  All
were running Spice 2G.6.



    Results

machine                 ckt A           ckt B           ckt C

VAX 11/785               65              320             189

masscomp 5400            50              278             172

Sun 3/160                50              277             172

MicroVax II              74              410             272

    Two of the three tests were run on a masscomp 5700, with
improvements of about 5%.

    With the exception of the MicroVax execution time, results obtained
by the simulations seemed to be identical in all substantive ways.
Because of the peculiarity of DEC floating point, the vax results
contained negative zero values, while the motorola implementation could
not.  The non-convergent case failed to converge in exactly the same
way, at the same time value, for all machines, giving some assurance
that numerical results are the same, even for ill-conditioned computations.