Approximate Computing with Stochastic Transistors' Voltage Over-scaling

Abstract

Approximate Computing with Stochastic Transistors’ Voltage Over-scaling Introduction Approximate computing is a promising technique for error resilient applications. The intrinsic variability of the transistor (reflectedonVTH)is a concern for traditional designs. Thisworkaddresses the variation as the source of performance shaping in approximate computing.

StochasticTransistorModel

The physical variations are summed up into threshold voltage(VTH)variability; The variation is modeled by adding a thermal noise to the gate voltage.

The added variability ensures Enoughdata points within a single transient simulation; The full spectrum of the Gaussian distribution is captured.

Approximate Computing The worst-case scenario is guaranteed by assigning the obtained value “x” to the counterpart of its correct value. Voltageover-scaling scheme explores the designs pace of approximate full adder (FA) up to 16-bit. Designconsideration includes: Technology node Operating frequency Energy and delay Process corner Temperature Quantification on Approximate Adder Mean Error Distance (MED) describes the actual value of error. Mean Relative Error Distance (MRED) describes the deviation from the expected value. Image Compression Using Approximate Adder saves up to 90% energy while preserving relative quality. Conclusion The IoT (Internet of Things) operations which consist of numerous error resilient applications can benefit from this work. In conclusion, it Embraces and models the variability of the transistor; Adopts the inherent stochasticity in approximate computing; Provides design space and improved energy efficiency References R. Li, R. Naous, H. Fariborzi, and K. N. Salama, “Approximate Computing with Stochastic Transistors’ Voltage Over-scaling,” IEEE Access, 2018. DOI: 10.1109/ACCESS.2018.2889747 P. Weckx, et al, “Defect-centric perspective of combined BTI and RTN time-dependent variability,” in 2015 IEEE International Integrated Reliability Workshop (IIRW), pp. 21–28.