On the Efficiency of Early Bird Sampling (EBS) an Error Detection-Correction Scheme for Data-Driven Voltage Over-Scaling

Abstract

An efficient implementation of voltage over-scaling policies for ultra-low power ICs passes through the design of on-chip Error Detection and Correction (EDC) mechanisms that can provide continuous feedback about the health of the circuit. The key components of a EDC architecture are embedded timing sensors that check the compliance of timing constraints at run-time and drives the computation to safely evolve toward the minimum energy point. While most of the existing EDC solutions, e.g., Razor [1], have proved hardly applicable to circuits other than pipelined processors, our recent work [2] introduced a lightweight EDC alternative for general sequential circuits, what we called Early Bird Sampling (EBS). As a key strength, EBS reduces the design overhead by means of a dynamic short path padding that alleviates the overhead of timing sensors placement. Moreover, EBS implements an error correction mechanism based on local logicmasking, a technique that is well suited for digital IPs w/o an instructionset. These features make EBS a viable solution to devise Data-Driven Voltage Over-Scaling (DD-VOS) for error-resilient applications. Aim of this work is to recap the EBS strategy and quantify its figures of merit under different power management scenarios. We thereby provide accurate overhead assessment for different benchmarks and run under different DD-VOS policies. Comparison against a state-of-art EDC scheme, i.e., Razor, demonstrates EBS shows affordable area penalty (3.6% against 71.6% of Razor), still improving the efficiency of DD-VOS. Indeed, EBS leads circuits through lower energy-per-operation (savings w.r.t. Razor range from 36.2% to 40.2%) at negligible performance loss, from 2% to 5% (as much as Razor).