Asymmetric multi-cores are the state of the art and the way to go if you want maximum throughput per silicon area and per Watt.
Modern PC CPUs have evolved away from providing the fastest possible execution times (K6 3-cycle IMUL ftw!), toward extracting parallelism from everyday code, to use their relatively few execution units as effectively as possible, especially when executing code that engineers hate.
OTOH if you want to just crunch through large amounts of data quickly, you need many execution resources and a way to keep them fed, but the silicon overhead for clever management turns into a problem rather than a strength. The transistors it uses could have been spent for execution, to greater effect.
Sure you can use a PC CPU to do number crunching, but you're not going to win against a custom-designed bunch of execution units that gets to use the same silicon area and fab technology. The only problem here is that you might need a more general PC CPU to drive it or otherwise, if you want it to be able to manage itself, you're compromising the design in the direction you wanted to avoid.
So in practice, you want the characteristics of both processor types.
You want at least one somewhat normal CPU core in a system to do housekeeping and to orchestrate all these little critters, but slapping together multiple large and branchy-code-friendly CPUs is wasteful.
In a closed-box single-task, single-user system, single-thread performance isn't as much of a priority as it would be on a PC, and ~same performance profile/thread portability between execution cores is completely irrelevant. Thus embedded can take leaps the PC market cannot, as soon as it makes sense.
Cell builds heavily on IBM's preexisting PowerPC ISA. The project might have been expensive and long in the making, but I don't perceive it as risky or poorly thought out. The overall idea is pretty sound, the PPC building blocks have existed and have proven themselves in practice, and it has been understood for a good while that other approaches to increased CPU performance have certain issues that may be acceptable for a PC that needs to run legacy software as fast as possible, but aren't such hot ideas for a system designed off a clean slate. The goals were pretty clear. It took so long to make it just right, not to get it finished at all in whatever form.
Oh and re RSX, I don't think it's a poor choice, it's a pretty fast GPU and a reasonable choice for the time frame where a decision had to be reached, it's just less impressive and IMO less elegant than the Cell BE. Cell is somewhat of an engineer's wet dream come true. It's pure and right.
