Resumen de Synthesis of variability-tolerant circuits with adaptive clocking

Alberto Moreno Vega

• Improvements in circuit manufacturing have allowed, along the years, increasingly com- plex designs. This has been enabled by the miniaturization that circuit components have undergone. But, in recent years, this scaling has shown decreasing benefits as we approach fundamental limits. Furthermore, the decrease in size is nowadays producing an increase in variability: unpredictable differences and changes in the behavior of com- ponents. Historically, this has been addressed by establishing guardband margins at the design stage. Nonetheless, as variability grows, the amount of pessimism introduced by these margins is taking an ever-increasing cost on performance and power consumption.

  In recent years, several approaches have been proposed to lower the impact of vari- ability and reduce margins. One such technique is the substitution of a classical PLL clock by a Ring Oscillator Clock. The design of the Ring Oscillator Clock is done in such a way that its variability is highly correlated to that of the circuit. One of the contributions of this thesis is in the automatic design of such circuits. In particular, we propose a novel method to design digital delay lines with variability-tracking properties.

  Those designs are also suitable for other purposes, such as bundled-data circuits or per- formance monitors. The advantage of the proposed technique is based on the exclusive use of cells from a standard cell library, which lowers the design cost and complexity.

  The other focus of this thesis is on state encoding for asynchronous controllers. One of the main properties of asynchronous circuits is their ability to, implicitly, work under variable conditions. In the near future, this advantage might increase the relevance of this class of circuits. One of the hardest stages for the synthesis of these circuits is the state encoding. This thesis presents a SAT-based algorithm for solving the state encoding at the state level. It is shown, by means of a comprehensive benchmark suite, that results obtained by this technique improve significantly compared to results from similar approaches.

  Nonetheless, the main limitation of techniques at the state level is the state explosion problem, to which the sequential modeling of concurrency is often subject to. The last contribution of this thesis is a method to process asynchronous circuits in order to allow the use of state-based techniques for large instances. In particular, the process is divided into three stages: projection, signal insertion and re-composition. In the projection step, the behavior of the controller is simplified until the signal insertion can be performed by state-based techniques. Afterwards, the re-composition generalizes the insertion of the signal into the original controller. Experimental results show that this process enables the resolution of large controllers, in the order of 10 6 states, by state-based techniques.

  At the same time, only a minor impact in solution quality is observed, preserving one of the main advantages for state-based approaches.