Near-Threshold Computing: Past, Present, and Future.

Abstract

Transistor threshold voltages have stagnated in recent years, deviating from constant-voltage scaling theory and directly limiting supply voltage scaling. To overcome the resulting energy and power dissipation barriers, energy efficiency can be improved through aggressive voltage scaling, and there has been increased interest in operating at near-threshold computing (NTC) supply voltages. In this region sizable energy gains are achieved with moderate performance loss, some of which can be regained through parallelism.

This thesis first provides a methodical definition of how near to threshold is "near threshold" and continues with an in-depth examination of NTC across past, present, and future CMOS technologies. By systematically defining near-threshold, the trends and tradeoffs are analyzed, lending insight in how best to design and optimize near-threshold systems.

NTC works best for technologies that feature good circuit delay scalability, therefore technologies without strong short-channel effects. Early planar technologies (prior to 90nm or so) featured good circuit scalability (8x energy gains), but lacked area in which to add cores for parallelization. Recent planar nodes (32nm – 20nm) feature more area for cores but suffer from poor delay scalability, and so are not well-suited for NTC (4x energy gains).

The switch to FinFET CMOS technology allows for a return to strong voltage scalability (8x gain), reversing trends seen in planar technologies, while dark silicon has created an opportunity to add cores for parallelization. Improved FinFET voltage scalability even allows for latency reduction of a single task, as long as the task is sufficiently parallelizable (< 10% serial code).

Finally, we will look at a technique for fast voltage boosting, called Shortstop, in which a core's operating voltage is raised in 10s of cycles. Shortstop can be used to quickly respond to single-threaded performance demands of a near-threshold system by leveraging the innate parasitic inductance of a dedicated dirty supply rail, further improving energy efficiency. The technique is demonstrated in a wirebond implementation and is able to boost a core up to 1.8x faster than a header-based approach, while reducing supply droop by 2-7x. An improved flip-chip architecture is also proposed.