Near Threshold Computing: From Single Core to Many-Core Energy EfficientArchitectures.

[摘要] Over the past four decades, the number of transistors on a chip hasincreased exponentially in accordance with Moore;;s law.This hasled to progress in diversified computing applications, such as healthcare, education, security, and communications. A number of societalprojections and industrial roadmaps are driven by the expectation thatthese rates of improvement will continue, but the impediments togrowth are more formidable today than ever before. The largest ofthese barriers is related to energy and power dissipation, and it isnot an exaggeration to state that developing energy-efficientsolutions is critical to the survival of the semiconductorindustry. Extensions of today;;s solutions can only go so far, andwithout improvements in energy efficiency they are in danger ofrunning out of steam.When examining the history of computers, a pattern emerges: successive generations of technologies, ranging from vacuum tubes to bipolar toNMOS-based technologies, were replaced when theirenergy overheads became prohibitive. However, there is no clearsuccessor to todays technology, CMOS. The available alternatives are far from beingcommercially viable, and none has gained sufficient traction, orprovided the economic justification for overthrowing the largeinvestments made in today;;s CMOS-based infrastructure. Therefore, there isa strong case supporting the position that solutions to the powerconundrum must come from enhanced devices, design styles andarchitectures, rather than a reliance on the promise of radically newtechnologies becoming commercially viable. This dissertation proposes that the solutionto this energy crisis is the universal application of aggressive lowvoltage operation across all computation platforms. This can beaccomplished by targeting so-called ;;near-threshold computing;; (NTC)and byproposing novel methods to overcome the barriers that havehistorically relegated ultra-low voltage operation to niche markets. In particular, this dissertation explores the performancebarrier that prevents thewidespread adoption of NTC and provides architectures for single-core,multi-core, many-core, and digital signal processing systems. As afinal proof of concept a 3D integrated prototype of 128 ARM Cortex-M3cores and 256MB of DRAM is presented that shows a 6.4X improvement inenergy-efficiency over the Intel Atom processor.

[发布机构] University of Michigan

[效力级别] Computer Science

[关键词] Low Power Architectures;Computer Science;Engineering;Computer Science & Engineering