Approximate Computing: Contributions to the Design of Arithmetic Circuits and Instruction-Set Architectures

Abstract

The growing demand for computing power, coupled with the limitations of the end of the Dennard scale, has challenged designers to find alternative solutions to maintain performance within energy and cost limits. Approximate computing (AC) has emerged as a promising approach to balance performance and energy efficiency in error-tolerant applications. However, many AC techniques focus on specific problems or require much intervention from the programmer. This work identified gaps that were transformed into research opportunities. One is related to approximate arithmetic circuits, which focus on single-bit operations, limiting the analysis of these circuits' physical behavior, accuracy, and performance on real platforms with larger inputs and outputs. There are also limitations in the loop perforation technique since once the perforation degree (pd) is established, the application metrics will improve only at the cost of accuracy. Adopting a strategy in which pd can use approximate hardware resources would overcome this limitation, and greater flexibility would be obtained without forcing additional compilation steps. There is little exploration of the use of approximate instructions, especially in the context of floating point operations, leaving an implementation gap that can be solved by introducing an additional level of approximation, replacing precise (non-approximate) instructions with approximate instructions, thus offering a hardware-level approximate technique over a source code that is already (or not) approximated by a software-level technique. Thus, this work aims at design space exploration (DSE) at different levels of abstraction, investigating the impact of AC on approximate arithmetic circuits, approximate instructions, loop perforation techniques, and approximate mathematical functions. In addition, extensions of the RISC-V architecture instruction set with support for AC are designed. AC techniques were integrated into a widely used platforms, such as SPIKE and ACCEPT, to provide a flexible and efficient infrastructure for developing of approximate systems. The results demonstrate that AC can significantly improve performance and energy efficiency without substantially compromising the accuracy of the systems. This work contributes to new approximate instructions, arithmetic circuits, and an energy model for approximate instructions. It explores the feasibility of these techniques in mathematical functions and control structures (loop) in applications that demand high performance but tolerate controlled errors.