Analysis and design on low-power multi-Gb/s serial links

Graduation date: 2012High speed serial links are critical components for addressing the growing demand for I/O bandwidth in next-generation computing applications, such as many-core systems, backplane and optical data communications. Due to continued process scaling and circuit innovations, today's CMOS serial link transceivers can achieve tens of Gb/s per pin. However, most of their reported power efficiency improves much slower than the rise of data rate. Therefore, aggregate I/O power is increasing and will exceed the power budget if the trend for more off-chip bandwidth is sustained.\ud In this work, a system level statistical analysis of serial links is first described, and compares the link performance of Non-Return-to-Zero (2-PAM) with higher-order modulation (duobinary) signaling schemes. This method enables fast and accurate BER distribution simulation of serial link transceivers that include channel and circuit imperfections, such as finite pulse rise/fall time, duty cycle variation, and both receiver and transmitter forwarded-clock jitter.\ud \ud Second, in order to address link power efficiency, two test chips have been implemented. The first one describes a quad-lane, 6.4-7.2 Gb/s serial link receiver prototype using a forwarded clock architecture. A novel phase deskew scheme using injection-locked ring oscillators (ILRO) is proposed that achieves greater than one UI of phase shift for multiple clock phases, eliminating phase rotation and interpolation required in conventional architectures. Each receiver, optimized for power efficiency, consists of a low-power linear equalizer, four offset-cancelled quantizers for 1:4 demultiplexing, and an injection-locked ring oscillator coupled to a low-voltage swing, global clock distribution. Measurement results show a 6.4-7.2Gb/s data rate with BER < 10⁻¹² across 14 cm of PCB, and an 8Gb/s data rate through 4cm of PCB. Designed in a 1.2V, 90nm CMOS process, the ILRO achieves a wide tuning range from 1.6-2.6GHz. The total area of each receiver is 0.0174mm², resulting in a measured power efficiency of 0.6mW/Gb/s. \ud \ud Improving upon the first test chip, a second test chip for 8Gb/s forwarded clock serial link receivers exploits a low-power super-harmonic injection-locked ring oscillator for symmetric multi-phase local clock generation and deskewing. Further power reduction is achieved by designing most of the receiver circuits in the near-threshold region (0.6V supply), with the exception of only the global clock buffer, test buffers and synthesized digital test circuits at nominal 1V supply. At the architectural level, a 1:10 direct demultiplexing rate is chosen to achieve low supply operation by exploiting high-parallelism. Fabricated in 65nm CMOS technology, two receiver prototypes are integrated in this test chip, one without and the other with front-end boot-strapped S/Hs. Including the amortized power of global clock distribution, the proposed serial link receivers consume 1.3mW and 2mW respectively at 8Gb/s input data rate, achieving a power efficiency of 0.163mW/Gb/s and 0.25mW/Gb/s. Measurement results show both receivers achieve BER < 10⁻¹² across a 20-cm FR4 PCB channel