At a recent MNFIG meeting the MACH II version of FORTH was demonstrated
on an Macintosh. It is subroutine threaded, many primitives
are assembled inline.  Consequently it is fast and big compared to 
conventional indirect threaded FORTH's.  Tho execution speed is not
the only criteria to compare FORTH implementations it is often cited
when comparing languages etc. I decided to dig out some benchmarks 
to get some basis for comparing execution speed.  I used the prime number
Sieve of Eratosthenes as published in Byte Jan 1983 pg 286. I used it
as close as possible to the way it was published making modifications only
as necessary. (Forth 83 variables, memory map limitations etc) I did not 
replace DROP DROP with 2DROP for example.
   The results are compiled from several sources and notes on my 
findings that explain a few items are in a seperate bulletin.
Fred Olson 
computer       Language              clock   time  notes
IBM PC   MMSFORTH ver 2.1            4.77    55.9  1
IBM PC   MVP PADS                    4.77    55.5  1
IBM PC   polyFORTH                   4.77    55.6  1
IBM PC   LMI PC/FORTH                4.77    55.3  1
IBM PC   LMI PC/FORTH+               4.77    97.2  1,5
IBM PC   QUEST FORTH-32              4.77   131.2  1,6
IBM PC   MS FORTRAN ver 3.2          4.77    23.7  1
IBM PC   BASICA (interpreted)        4.77  2002.1  1
Compaq   HS-FORTH                    4.77?   48    2
PC Clone F83-MSDOS                   4.77    69.   2,4
PC Clone LMI ver 1.2 (fig)           4.77    56.   2,4
PC Clone F83-MSDOS                   8.00    41.   2
PC Clone LMI ver 1.2 (fig)           8.00    31.   2
Compaq 386 F83-MSDOS                16.00     9.2  2
Osborne  F83-CP/M                    4.00    94.   2
Apple ][ F83X (Wil Baden)           ~1.00   191.   2
Hitachi 6303 fig-FORTH                .61   219.   2
TERAK PDP 11-03  poly FORTH          ?       75.   2
PDP 11-60 (Scanner Lab) poly FORTH   ?       12.5  2
Macintosh  MACH II                  ~8.      12.3  2
68000      polyFORTH                 8.00    29.0  3
68000      Assembler                 8.00     4.9  3
VAX 780    "C"                       ?        1.4  3
NC4000P    novixFORTH (cells)        6.00      .85 3
NC4000P    novixFORTH (optimised)    6.00      .45 3
IBM 3033   PL/1                      ?         .36 3

NOTES:
1 From Pearson & Gerow , "Comparison of FORTH's for the 
  IBM PC" , a document from the Boston FIG chapter dated 3-OCT-84 that was
  distributed with the FIG chapter mailings. It compares several FORTH's on
  several aspects including the Sieve timings.
  version               vendor (see ads in Forth Dimensions)
  MMSFORTH ver 2.1     Miller Microcomputer Systems
  MVP PADS             Mountain View Press
  polyFORTH            Forth Inc.
  LMI PC/FORTH         Laboratory Microsystems Inc.
  LMI PC/FORTH+        "           "           "
  QUEST FORTH-32       Quest Research Inc, Huntsville, AL (F.D. Vol 6 #1 pg 32)
  
2 These are timings I did in April 1987.
  LMI ver 1.2 (fig)   see above.  Old version (~1982) that I have on loan.
  F83-MSDOS  Henry Laxen and Mike Perry's public domain 83 Standard FORTH.
  F83-CP/M    "     "                 
  F83X for Apple][ similar to F83. Implemented by Wil Baden / Orange Cty CA FIG
  6303 fig-FORTH  adapted from 6800 figFORTH by Peopleware Systems, Mpls,MN
  MACH II for Macintosh by Palo Alto Shipping Co. Menlo Park, CA 

3 These timings came from a sheet distributed by NOVIX that has the COMSOL
  logo.  COMSOL is Computer Solutions Limited, Surrey,England. These timings 
  were probably done in 1986. 
4 The difference in the time for the LMI fig and the F83 caught my attention
  and I looked for an explaination. Both are conventional 64K ITC implemen-
  tations. In fact both implement the virtual machine using the same registers.
  Using DEBUG I found that in LMI's FORTH some code words in the kernel have
  NEXT inline rather than jumping to it. (I have not determined how many or
  which words. The assembler's "NEXT" generates jumps to the NEXT code. Also
  note that this is a rather old version.)  To verify that this accounted
  for the difference in timing I patched F83's centralized NEXT to jump to
  a temporary copy and did the Sieve again. It took 83 seconds, 14 seconds
  longer. If the 14 seconds taken by the extra jump each time thru NEXT is
  subtracted from F83's normal time to get an estimate of F83's speed if
  NEXT were compiled in-line, the resulting 55 seconds corresponds closely
  to the other FORTH's on the PC.  Note that F83's centralized NEXT allows
  the implementation of the high level debugger, an extremely useful tool.
5 PC/FORTH+ is a 32 bit implementation from LMI which according to their 
  recent catalog has "speed penalty ... in the range of 30-50% for the 8086
  processor family."
6 QUEST's FORTH-32 is (was?) a 32 bit implementation.

Fred Olson