For mid-tier computation I'd (inspired by the PS3) include several microprocessors connected via FIFO channels. By cost-cutting these cores I should be able to fit more in! Besides, FIFO channels are an easier way to manage concurrency!
Include firmware to let you program these.
And for the low-end I'd be tempted to include a Reduceron to coordinate all the other cores. Even if it'd be poorly suited to doing any of the calculation itself, it could be easier to program!
2/3?