Energy conservation in asynchronous systems using self-adaptive fine-grain voltage scaling

Dynamic voltage scaling is a technique for lowering energy consumption in microelectronic systems when the performance required is at times lower than the system's peak capability. This paper proposes a new voltage scaling technique for asynchronous systems which allows fine-grain scaling of supply voltage, i.e. each stage in the pipeline is allowed to scale its voltage independently. Instead of applying a single scaled supply voltage to all the stages in a pipeline, each stage operates at a voltage which is just sufficient for it to run at a throughput supported by the pipeline. This results in higher energy savings as stages with smaller latency can now run at a lower voltage than the voltage required by higher latency stages. Voltage scaling for each stage is guided by local `congestion-starvation' information generated only from its handshake signals, not requiring any global coordination. The technique described in this paper is self-adaptive, on-chip, and allows continuous scaling of voltage levels. It allows choosing appropriate level of voltage scaling granularity with desired energy-area trade-offs for a given system. Fine-grain approach has the advantage of more energy savings while it has higher area overhead as compared to a coarse-grain approach. The proposed technique is applicable to linear pipelined systems, as well as systems with more complex topologies, including pipeline constructs for parallel, conditional, and repetitive computation. We present hardware designs of voltage controller and voltage regulator used in our technique. We simulated our technique at behavioral level in Verilog for five algorithms modeled as micropipelines and simulation results show that our technique achieves upto 85% dynamic energy savings with a throughput degradation of less than 10%.