Addressing the energy challenges for modern data science

With emerging sustainability goals and the proliferation of the use of data and associated AI analysis, the energy consumption of computing based processing is becoming an increasingly important concern. Research into lower power design techniques to reduce the energy consumption of computing systems has been a topic of research for over three decades. The paper starts with a discussion of the earlier low power research on reducing dynamic power consumption, undertaken in the UK in the 1990s by researchers including Steve Furber. In particular, we concentrate on research undertaken on the Powerpack and PREST projects by the universities of Sheffield, Manchester, Liverpool and Queenʹs University Belfast. The work focused on not only developing techniques to reduce dynamic power consumption but also on demonstrating the approaches in a real working case study, in this case, Mitel’s Viterbi decoder design. The paper then covers some of the emerging approaches in addressing the current energy challenges of computing systems that will deliver our data inspired, artificial intelligence processing needs. In particular, it examines the approach of approximate or transprecision computing as a means to improve hardware performance by relaxing accuracy metrics.