This paper presents a new technique for global energy optimization through coordinated functional partitioning and speed selection for embedded processors interconnected by a high-speed serial bus. Many such serial interfaces are capable of operating at multiple speeds and can open up a new dimension of trade-offs to complement today's CPU-centric voltage scaling techniques for processors. We propose a multi-dimensional dynamic programming formulation for energy-optimal functional partitioning with CPU/communication speed selection for a class of data-regular applications under performance constraints. We demonstrate the effectiveness of our optimization techniques with an image processing application mapped onto a multi-processor architecture with a multi-speed Ethernet.