Low-Latency Energy-Recovery Circuitry.

[摘要] Voltage scaling has diminished with each advancement in process technologies, making power dissipation one of the primary design considerations for modern digital systems across all market segments. This dissertation describes a novel charge-recovery logic family and a resonant-clocked dynamic logic that utilize energy-recovery techniques to recycle charge from the system, effectively reducing dynamic power dissipation.We propose a novel charge-recovery logic, called Enhanced Boost Logic, which achieves high efficiency and high performance operation through the use of aggressive voltage scaling, gate overdrive, and charge-recovery techniques. Fabricated using a 0.13μm technology with a 3nH on-chip inductor, a 14-tap 8-bit FIR filter implemented using Enhanced Boost Logic achieves operating frequency up to 600MHz with only 1.5 cycles of latency overhead compared to a static CMOS implementation. At its natural frequency of 466MHz, the FIR dissipates 39.1mW. With a figure of merit equal to 93.6nW/MHZ/Tap/InBit/CoeffBit, it achieves 29% improvement compared to previously reported FIR filter test-chips with equal or greater sampling rate than our design at its 466MHz resonant point.We also propose a dynamic logic family, call Dynamic Evaluation Static Latch, synchronized by a two-phase resonant clock, which achieves dynamic-logic levels performance with significant power reduction. Fabricated in a 90nm technology with an 0.41nH integrated inductor, a resonant-clocked FPU implemented using this methodology achieves clock speeds up to 2.07GHz. At its resonant frequency of 1.81GHz, it dissipates 334mW, yielding 31.5% lower power and 32% more GFLOPS/W over a conventionally-clocked version of the same FPU implemented on the same die. Relying on circuit, logic, and architectural optimizations, the resonant-clock FPU breaks new ground along several metrics. Specifically, with a total area of 0.391mm2 including the on-chip inductor, it occupies the smallest footprint among competing stat-of-the-art reduced-latency FPUs. Moreover with an overall latency of 64 FO4, it achieves the shortest overall latency among state-of-the-art reduced-latency FPUs. Delivering 10.82GFLOPS/W, this resonant-clock FPU achieves the highest energy efficiency among state-of-the-art continuously data-streaming FPUs.

[发布机构] University of Michigan

[效力级别] Resonant Clocking

[关键词] Charge-recovery Logic;Resonant Clocking;Low-power VLSI;Electrical Engineering;Engineering;Electrical Engineering