Exploiting FDSOI Towards Minimum Energy Point Operation in Processors and Machine Learning Accelerators

Abstract

The energy consumed by a processor to execute a certain task, depends strongly on several run-time conditions, such as the current workload, speed requirements, required computational precision, temperature, … For each possible scenario, it is possible to find a minimum energy operating point. In FDSOI technologies, this MEP is not only be influenced by the chip's supply voltage, but also by the technology's threshold voltage, tunable through the back bias. This increases the MEP exploration space, rendering much more optimal trade-offs.