High performance computing is critical for the needs of scientific discovery and economic competitiveness. An extreme-scale computing system at 1000x the performance of todayβs petaflop machines will exhibit massive parallelism on multiple vertical fronts, from thousands of computational units on a single processor to thousands of processors in a single data center. To facilitate such a massively-parallel extreme-scale computing, a key challenge is power. The challenge is not power associated with base computation but rather the problem of transporting data from one chip to another at high enough rates. This thesis presents architectures and techniques to achieve low power and area footprint while achieving high data rates in a dense very-short reach (VSR) chip-to-chip (C2C) communication network. High-speed serial communication operating at ultra-low supplies improves the energy-efficiency and lowers the power envelop of a system doing an exaflop of loops. One focus area of this thesis is clock synthesis for such energy-efficient interconnect applications operating at high speeds and ultra-low supplies. A sub-integer clockfrequency synthesizer is presented that incorporates a multi-phase injection-locked ring-oscillator-based prescaler for operation at an ultra-low supply voltage of 0.5V, phase-switching based programmable division for sub-integer clock-frequency synthesis, and automatic calibration to ensure injection lock. A record speed of 9GHz has been demonstrated at 0.5V in 45nm SOI CMOS. It consumes 3.5mW of power at 9.12GHz and 0.05ππ2 of area, while showing an output phase noise of -100dBc/Hz at 1MHz offset and RMS jitter of 325fs; it achieves a net πΉπππ΄ of -186.5 in a 45-nm SOI CMOS process. This thesis also describes a receiver with a reference-less clocking architecture for high-density VSR-C2C links. This architecture simplifies clock-tree planning in dense extreme-scaling computing environments and has high-bandwidth CDR to enable SSC for suppressing EMI and to mitigate TX jitter requirements. It features clock-less DFE and a high-bandwidth CDR based on master-slave ILOs for phase generation/rotation. The RX is implemented in 14nm CMOS and characterized at 19Gb/s. It is 1.5x faster that previous reference-less embedded-oscillator based designs with greater than 100MHz jitter tolerance bandwidth and recovers error-free data over VSR-C2C channels. It achieves a power-efficiency of 2.9pJ/b while recovering error-free data (BER< 10β12) across a 15dB loss channel. The jitter tolerance BW of the receiver is > 200MHz and the INL of the ILO-based phase-rotator (32- Steps/UI) is <1-LSB. Lastly, this thesis develops a time-domain delay-based modeling of injection locking to describe injection-locking phenomena in nonharmonic oscillators. The model is used to predict the locking bandwidth, and the locking dynamics of the locked oscillator. The model predictions are verified against simulations and measurements of a four-stage differential ring oscillator. The model is further used to predict the injection-locking behavior of a single-ended CMOS inverter based ring oscillator, the lock range of a multi-phase injection-locked ring-oscillator-based prescaler, as well as the dynamics of tracking injection phase perturbations in injection-locked masterslave oscillators; demonstrating its versatility in application to any nonharmonic oscillator.
Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/D8H437M5 |
Date | January 2018 |
Creators | Gangasani, Gautam |
Source Sets | Columbia University |
Language | English |
Detected Language | English |
Type | Theses |
Page generated in 0.002 seconds