This page will list Total Impact G3/MP card comments since they are often are too long for the main news page. If you have Total Impact G3/MP info to share, please contact
me.
"
Mike
I read your comments on the Total Impact with interest. It is true that in order to use the cards effectively you probably have to modify your app, but the API generally follows the official Mac MP API. The catch is that the cards use a distributed memory model, while the Mac assumes a shared memory model. For some applications this can be a significant impediment, for others it may be a minor tweak.
The distributed memory architecture is also what allows the G3 to be used effectively. The G3 lacks instructions for synchronizing shared memory, which does make them nearly useless under the current Mac MP standard (or most SMP architectures for that matter).
On the bright side, the PR statement that "this kind of power hasn't been available on the desktop" is a major understatement. A Mac with 40 G3s is effectively a supercomputer, and can compete with systems costing several hundred thousand dollars.
The heat is a real problem - but the power requirement is a more direct
problem, and the boards require two slots each because of their physical
dimensions. You could never run 6 boards in a Mac, even if it had enough
slots, but the cards do work fine in an expansion rack. Total Impact
announced a custom rack which they state can hold up to 10 TI boards, for a
total of 40 CPUs. The stock 14 slot SBS Bit 3 PCI expansion rack only has
enough power for 3-4 cards itself, but the racks can be chained (there is a
limit somewhere that I don't recall, but it's pretty high).
As far as supporting existing software, I'm not the expert here but according
to TI's information that is done. As I qualified in my earlier statement, it
may not be terribly effective. It is possible because the Apple MP libraries
are shared. TI simply replaces the Apple libraries with their own, and
applications are dynamicly linked to the new calls. I suspect that
applications which rely on shared memory will either run very slowly or not at
all (I have no idea how many MP applications make this assumption). Software
should really be designed to take advantage of TI's distributed architecture.
The beauty of this design is that there should be little or no disadvantage
when running on current Apple/Daystar MP boxes, so there is still only one
software version required.
J. Karl Armstrong
"
