Abstract:
Approximate computing targets the intrinsic error resilience of compute-intensive
applications for gaining computing efficiency with minimal degradation in the output quality. Arithmetic units constitute the most resources in compute-intensive
applications. Among the basic arithmetic units, divider exhibit the highest complexity, therefore, the design of resource-efficient approximate dividers is needed.
In this thesis, we propose a unified approximate multiplier and divider (INAxD)
design. The proposed design used the lookup-table to determine the approximate
divisor-reciprocal. The divisor-reciprocal was then multiplied with the dividend to
acquire the quotient. Furthermore, to provide a trade-off between resource efficiency
and output quality, the design was rendered error-configurable by the number of approximate divisor-reciprocal bits to be acquired from the lookup-table. Error and
hardware characteristics of the proposed design was evaluated by comparing with the
accurate restoring divider, Radix-2 SRT divider, Goldschmidt divider and NewtonRaphson divider. For N = 8, the INAxD offered upto to 1.10×, 1.56×−1.89×,
2.67× energy improvement for >0.3% error bias when compared with the accurate
dividers (Goldschmidt and Newton-Raphson). The power consumption of the proposed design was also optimal when configurability factor was involved. Finally,
we performed application-level evaluation showing that the proposed design enable
significant power and area reduction with minimal degradation in the output quality
