At one of the Motorola developer conferences a third-party was showing off a multi-cpu system that can be upgraded by adding CPUs via PCI card.
The PCI slots were upgraded with Rapid IO ports, so the extra CPUs had direct access to the main processor fabric switch.
But without the PCI slots being upgraded with a RIO or HT port, the PCI slot bandwidth is way too slow for true SMP using this method.
You can probably offload tasks to independent computers (which can be on a PCI slot), but then you might as well just set up a clustering system.