[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Number Cruncher - more details
I got some mails asking for more info on the Number Cruncher (NC) card.
Although the description "it's a FPE that works" is pretty good, I've
gathered some more details about it.
Please excuse that I don't have replied to all who sent me letters -
instead I wrote the stuff below.
The NC is a _math_coprocessor_ for the Apple II. (Entire family - as
far as I understood.) It fits into one of your slots and is compatible
with virtually anything else you can squeeze into your computer.
The core of the NC is a Motorola 68882 floating point unit, one of the
most sophisticated coprocs in this price class. Since it supports most
of the math functions that are located in the SANE toolset, it's an
excellent choice for a "hardware-based" math collection.
The 68882 features high precision math (with slightly less accuracy
when using transcedental functions, but nothing to worry about) and
less support for exception handling. Except for these two really minor
and unimportant facts, it offers more or less the same functionality.
BTW, the Mac has the same FPU (floating point unit), so this processor
is even "Apple approved".:)
The quality of the NC
The NC was built while the FPE (Floating Point Engine, a compatible but
no longer available math coprocessor) was still being sold. The reason
why the engineer saw a need for the NC: the FPE had some major design
flaws (incl. bugs in the GALs or how they are called) and did not
behave well.
That's why the hardware engineer decided to build a fully compatible
board. The AUGE, Europe's largest user group, helped him finance
prototypes etc. The result was a 100% FPE compatible coprocessor card
containing only the best parts available. All cards are "manu"factured
and tested.
Price
Expect a price of ca. US$200 without 68882. However, the exact cost of
the card is not known - it's only an assumption (rather a bit less).
Software support
Since the NC supports the same interface, you can use the FPE patches.
They accelerate most SANE tools by a yet unkown factor (see below). The
patches are freeware.
Also, some programs do exist that access the card(s) directly, not
going through the slow toolbox dispatchers. By pumping the values into
the NC/FPE and fetching the result immediately you're able to write
extremely fast routines.
Acceleration
It's pretty hard to give you a concrete number in the matter of
acceleration. It really depends on the kind of software you're using.
For example, don't expect a noticeable effect in a music program. When
working with AppleWorks GS' spreadsheet, your GS should be
significantly faster, though. And, in a math application that accesses
the FPE/NC directly you'll get a speed increase of ca. 300%. (Sorry, I
forgot the name of this program..:-)
Put briefly, you'll notice a "turbo-effect" sporadically, especially on
unwarped/unzipped systems.
Comparison of DSP/NC
I was asked which board is superior, the DSP card (coming sooner or
later..:) or the NC. There is a simple answer to this question: you
cannot compare the two cards!
The NC is a math coprocessor card. It will accelerate math intensive
tasks such as recalculating a spreadsheet or evaluating formulae. It
won't accelerate anything else.
The AIIdsp contains a Digital Signal Processor, a special CPU that is
specialized in time critical real-time tasks such as pattern/voice
recognition, sound processing, etc. This "class" of chips is one of the
worst for doing math stuff, but thanks to the very high clock rate of
20MHz it will eventually come close to the NC (though with far less
precision, namely 24bit instead of 80bit). We will ship or at least
intend to ship the AIIdsp with SANE patches because it _will_
accelerate math functions, so it's less important if this CPU has been
built for this purpose or not.
I have no idea how good the math routines for the DSP are - I think
it's safe to assume that the 68882 conforms much more to the SANE
standard (er, actually it bases on the IEEE 754 standard).
Conclusion: if you need a math accelerator, consider the NC. If you
need a DSP, wait for the AIIdsp. The AIIdsp will eventually overtake
some math routines, but the NC can do it better.
Final note: I'm not a hardware guy, so I know that there are millions
out there who know more the above. Please post corrections if anything
I have said is not correct. Thanks!
cu
henrik
--
BRIGHT SOFTWARE * P.O.Box 18 * 4153 Reinach 2 * Switzerland
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
voice (0)61 2619454 gudat@avalon.physik.unizh.ch Simple Solutions to
fax (0)61 7115263 gudath@ezinfo.vmsmail.ethz.ch Complex Problems.