Blue Gene isn't SMP. It is a fine-grained message passing architecture, kind of a modern remake of the old 1980s massively parallel supercomputers, with gobs of relatively simple processors. In a sense, it is basically a computing cluster that has been cut down to fit in a box, and with the same limitations on what it is useful for.
The poor multiprocessor scalability isn't intrinsic to the PPC, it was an explicit design choice. The Power series processors actually do have all the glue logic and memory hardware to scale extremely well in multiprocessor environments. These features were probably not included for two reasons. First, it was likely a cost saving measure by Apple and IBM, as vanilla SMP systems are much less costly to produce. Second, IBM may not want to cannibalize the sale of their Power systems. The additional features don't come cheap; with its huge cache, onboard memory management, and ccNUMA, the Opteron has a substantially larger die size than the PPC970 for the same fab process which translates into cost.
