🚀 go-pugleaf

>>>>> "Nathanael" == Nathanael J Henderson <nahender@plains.NoDak.edu> writes:
In article <DFAEK7.wDz@ns1.nodak.edu> nahender@plains.NoDak.edu (Nathanael J Henderson) writes:


    Nathanael> Clifford T. Matthews (ctm@ardi.com) wrote:

    Cliff> [Bias alert: I founded ARDI, the makers of Executor,
    Cliff> mentioned below]

    Cliff> For $49, students can buy the Executor/DOS+Executor/Linux
    Cliff> student combination.  Executor/DOS will run under DOS,
    Cliff> Windows '95, and OS/2 (and maybe Windows NT/Intel with the
    Cliff> right switches).  Executor/Linux

    Nathanael>    What exactly is the market for this thing, anyway?
    Nathanael> The point of buying a Mac is much more than the
    Nathanael> software available--it's the great interface and some
    Nathanael> of the fastest hardware available.  I suppose you could
    Nathanael> make a case for PC users 'stuck' in an all-Mac
    Nathanael> environment that need only the barest minimum of
    Nathanael> compatability/capability, but if that's the case you
    Nathanael> seem to have a very small market.

The market is anyone who would like to run Mac software on a PC, for
whatever reason.  This includes people who have a PC at work and a Mac
at home, people who have children in schools that use Macs and have
PCs at home, computer labs that have PCs that could be dual use, and
of course people with 68040 based Macs that are too slow.  Currently
we have many scientists using NIH Image, an image processing program
that is only available for the Mac, using Executor.  There is a lot of
"third tier" software that is Mac specific that people would like to
run on their PCs.  I don't think the market is particularly small, but
if you do, that doesn't bother me.  To each his own.

    Nathanael>    I'm also quite skeptical of the speed you give
    Nathanael> ("twice as fast as a 25Mhz 040") since *Apple's*
    Nathanael> emulator doesn't come close to that on equivilent
    Nathanael> hardware (to the 90 Mhz Pentium).

First let me apologize.  I have just reviewed our integer benchmarks
(reproduced below) and it looks like using them the speedup isn't a
factor of 2, but 1.65.  I probably was thinking of the first item in
the test, which is Speedometer's "CPU benchmark", where the speedup is
1.8 (close to 2).  At one time we were considering some modifications
that would bring those numbers up, but we decided the synthetic CPU
was fast enough and that it was more important to add functionality to
Executor and increase Executor's compatibility.

Even with that said, you *should* be skeptical.  After all, this is
usenet.  Not only do kooks populate usenet, but people who do serious
work are often reluctant to post at all because the net can be such a
tar baby.  So please don't take my word for it.  Pick up our white
paper, written by our synthetic CPU's author
(ftp://ftp.ardi.com/pub/SynPaper) and pick up a demo of our software
(available from the same FTP site) and try it out yourself.  Please
note, my original post clearly mentioned many of the incompatibilities
our total emulator has.  We're quite up front about them, although
they're not due to problems in our synthetic CPU; they're due to the
inherent difficulty of rewriting the MacOS and Toolbox from scratch.

    Nathanael> Have you castrated
    Nathanael> the emulator in order to make it simpler, or have you
    Nathanael> pulled off a tech-miracle?

Using your definition I would have to say tech-miracle.  We don't use
their emulator at all.  In fact, we use no Apple code at all.  Our
emulator uses dynamic recompilation into native x86 code.  Details are
provided in the paper I mention above.

    Nathanael> (Or are you just being a
    Nathanael> little...optimistic?)  Not that I doubt such speeds can
    Nathanael> be obtained--just not running normal programs.

Please bear in mind that the speeds I'm talking about are for the
compute intensive portions of the program.  The graphics portions will
vary depending on what graphic mode you're using and whether or not
Executor can linearly map your video board into memory.  But the
speeds I'm talking about are real and have been observed under real
programs like NIH Image as well as Speedometer.  We have done nothing
to artificially raise our speedometer speeds.

	--Cliff
	ctm@ardi.com

From SynPaper:

XII.  Benchmarks

These performance numbers were computed with Speedometer 3.23.
We've removed the floating point tests from the list since they do
not measure syn68k's speed.  Syn68k contains no special provisions
for Speedometer's benchmarks and we believe that these numbers are
indicative of syn68k's performance for many other CPU-intensive
programs.


		Quadra	Pentium	486DX4	486DX/2
                  610	 90MHz	 75MHz	 66MHz
		------	------	------	------
CPU		16.018	28.833	15.727	13.840

Dhrystones	19.586	21.886	12.084	 9.424
Tower		18.909	27.130	12.235	11.556
Quicksort	17.759	27.105	15.606	13.919
Bubble sort	18.409	31.154	19.286	16.875
Queens		19.083	38.167	19.083	18.320
Puzzle		22.083	44.167	23.661	21.032
Permutations	21.019	28.564	11.604	12.242
Int. Matrix	24.200	26.469	19.369	16.608
Sieve		23.362	60.290	33.982	30.145
		------	------	------	------
Average		20.490	33.881	18.582	16.680



Preliminary analysis suggests that we average a roughly 3:1
instruction count increase when translating to 80x86 code.  We have
not yet taken rigorous measurements, but the 3:1 figure is lent
some credence by the fact that our 75MHz 486DX4 gets nearly the
same performance as our Quadra 610.  We believe that inter-block
register allocation will noticeably improve this ratio.