Energy-aware clock-frequency assignment in microprocessors and memory devices for dynamic voltage scaling

Dynamic supply-voltage scaling (DVS) can reduce the energy consumption of microprocessors, but most DVS schemes only scale the clock frequency of the microprocessor and ignore the memory system. In this paper, we show how more energy can be saved by changing the clock frequency of the memory as well as that of the microprocessor in a coordinated fashion. The contributions of this paper include: 1) consideration of both the energy and access time of the memory; 2) derivation of a mathematical formulation of a system-wide energy model as a function of the clock frequencies of the microprocessor and memory; 3) derivation of analytic solutions of system-wide energy-optimal clock-frequency pairs for the microprocessor and the memory, and, finally, 4) extension of the frequency-assignment technique to handle discrete voltages and frequencies. Cycle-accurate system-level energy simulation shows that the proposed scheme can save up to 50% more energy than previous DVS schemes. Our approach can also be applied to other synchronous peripheral devices.