Abstract
Power dissipation has became a critical concern for present VLSI design in recent years. DPM (Dynamic Power Management) is a common methodology which can dynamically scale the power level of ASIC to adapt their requirements at runtime. Unfortunately, power management usually accompanies some performance degradation. So how to eliminate the unnecessary power dissipation with minimum harm of performance will become a significant challenge for designers. Many preceding researches of DPM just focus on complicated mathematical solutions, which are hard to implement in hardware. When DPM methodologies are practiced in pure software, their efficiencies highly rely on the operating system. Moreover, huge computation overhead of DPM manipulation can become a burden of operation system to diminish the ability of main processor. Hence, a hardware-based DVFS power manager is proposed. Our structure contains simple computations that can be easily implemented in hardware but still maintain a well power managed facility. AES (Advanced Encryption Standard) is our DPM object, which is a fast cryptological scheme. Because this device does not always need the peak performance, it provides chances to reduce its overall energy dissipation with an appropriate DPM methodology. Different from traditional DPM researches, our proposed methodology contains many practical concerns like the level transition overhand or the power transform efficiency. These crucial concerns make it closer to the reality, but the complexity of DPM is also more difficult than others. Addition to basic DPM research, we combine 3 novel strategies into our primary DPM to handle some special situations. After these exception handling, our final methodology can be more stronger than the primitive one. In order to demonstrate our DPM efficiency, we generate a serial of user-defined test patterns which contain a variety of different workload distribution. For finding a general best solution, we construct a SystemC model to experiment and exploit many different DPM policies. The experimental results show that our ideal energy reduction can achieve 59.3% by the offline methodology, and the practical online methodology can reduce 53.0% energy dissipation with just 6.0% performance degradation. With many practical concerns, our DPM methodology is still much close to the ideal offline results.