The 9500 did not have built in video.
Doh, I forgot about that. My faulty memory filled in that it had the same framebuffer as the 7/8xx's stashed on the board, but thinking about it at all I know that's wrong.
Just as an aside, the Developer Note that covers the 9600/200MP has this to say about it:
Dual Processor Configuration:
In the 9600/200MP model, two PowerPC 604e microprocessors are on the processor card.
With applications that support the new multiprocessor API, the MP configuration
provides up to 2 times the performance of the equivalent single-processor computers.
The operation of the dual-processor configuration is asymmetric multiprocessing. One
processor is the primary processor: it runs the Mac OS and handles interrupts from the
I/O systems. The second processor runs MP tasks as set up by the primary processor.
Unfortunately there's no documentation, not even a block diagram, in that document that hints at how they actually hang the multiple CPUs off the bus. Granted that's not the sort of information Apple used to give out about their machines, since they never intended for anyone to run alternate OSes on them. Therefore it's difficult to take much away from this statement, since the "Asymmetric" part may strictly refer to their particular software implementation rather than a limitation of the hardware.
Also, for shiznet and giggles I downloaded a random version of the Linux 2.4 kernel source. Here's what I found in arch/ppc/kernel/pmac_smp.c:
/*
* SMP support for power macintosh.
*
* We support both the old "powersurge" SMP architecture
* and the current Core99 (G4 PowerMac) machines.
*
* Support Macintosh G4 SMP by Troy Benjegerdes (hozer@drgw.net)
* and Ben Herrenschmidt <benh@kernel.crashing.org>.
*
* Support for DayStar quad CPU cards
* Copyright (C) XLR8, Inc. 1994-2000
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License
* as published by the Free Software Foundation; either version
* 2 of the License, or (at your option) any later version.
*/
The comments in that file are interesting reading. Among other things it seems to say that the CPUs in those machines are not mapped to Open Firmware device tree entries, you have to use some not-totally-reliable probing of registers in the Hammerhead memory controller to try to determine if they're present and what exact flavor of "PowerSurge" you're dealing with. The init code for the "Core99" machines is far cleaner with a lot less "cross your fingers and pray you got it right" moments. Given that I can see why they ditched the PowerSurge code in Linux, and it also probably explains why OS X has *no idea* about those extra CPUs either. (OS X is *very* Open Firmware-centric.)