Extending the POWER Architecture with Transprecision Co-Processors

Abstract

Approximate Computing is an emerging computer architecture paradigm which allows to trade off machine accuracy for energy efficiency. Initial studies have shown that Approximate Computing can significantly reduce the energy consumption of computer systems by introducing certain means of imprecision into the microarchitecture of the processor and memory units. However, it is not clear how quickly concepts that affect the precision will be adopted by enterprise class systems, because they need to deliver a high level of reliability, availability and serviceability (RAS). In this paper we propose to extend the traditional POWER architecture with closely coupled, transprecision co-processors. Transprecision co-processors use a re-configurable data path that can be optimized according to the precision requirements of the running algorithm. We show how FPGAs can be used to implement transprecision co-processors and conclude that CAPI is providing an innovative solution for coupling transprecision accelerators tightly to a CPU.