Next-Generation Energy-efficient Approximate Computing Systems Design
Data is driving the fourth Industrial Revolution. The Information Age combines technological innovation with the affordability of electronic systems.
Project leader
Dates
March 2017 - March 2019
Sponsors
Royal Society (International Exchange Scheme)
Description
Many emerging applications are providing new digital services. These applications use big data: continuous data acquisition, computation and storage systems. Examples include computer vision, data analytics and pervasive systems. But widespread adoption increases energy-related costs for technology and service providers. Data proliferation is pushing the limits with unprecedented processing requirements, challenging performance needs. In turn, this means that existing computing systems need costly arrangements. They need to provide large-scale parallel processing at higher performance levels.
Fabrication processes such as smaller devices or larger scale integration will be unable to meet this challenge. We need innovative thinking to provide scalable performance growth with a large reduction in energy costs. For example, consider how our brains achieve extremely high energy efficiency. This is despite the fact that they process images continuously in real-time. Our brains do this by ranking the importance of data. We then process data in smaller and parallel units. We compute data with high informational value in more detail than data with less significance.
This is the principle of adaptive approximate computing. In this project, we will explore this disruptive new paradigm. We will modulate data processing and movement based on significance.
We will deliver pioneering research to establish such a paradigm. We will achieve transformational reduction in energy consumption for emerging big data applications. To do this, we will design intelligent processors that can modulate data processing based on their significance. Higher level resources will process data deemed more significant. Insignificant or less significant data will be processed with imprecise logic, using fewer resources. This will reduce energy consumption.
We will incorporate key hardware and software innovations in the processor design. Hardware innovations will include novel approximate data processing and inference units. Software features will involve support for significance-modulated computing capabilities.