For decades, digital design has been primarily dominated by clocked circuits. With larger scales of integration made possible by improved semiconductor manufacturing techniques, relying on a clock signal to orchestrate logic operations across an entire chip became increasingly difficult. Motivated by this problem, designers are currently considering circuits which can operate without a clock. However, the wide acceptance of these circuits by the digital design community requires two ingredients: (i) a unified design methodology supported by widely available CAD tools, and (ii) a granularity of design techniques suitable for synthesizing large designs. Currently, there is no unified established design methodology to support the design and verification of these circuits. Moreover, the majority of clockless design techniques is conceived at circuit level, and is subsequently so fine-grain, that their application to large designs can have unacceptable area costs. Given these considerations, this dissertation presents a new clockless technique, called self-resetting stage logic (SRSL), in which the computation of a block is reset periodically from within the block itself. SRSL is used as a building block for three coarse-grain pipelining techniques: (i) Stage-controlled self-resetting stage logic (S-SRSL) Pipelines: In these pipelines, the control of the communication between stages is performed locally between each pair of stages. This communication is performed in a uni-directional manner in order to simplify its implementation. (ii) Pipeline-controlled self-resetting stage logic (P-SRSL) Pipelines: In these pipelines, the communication between each pair of stages in the pipeline is driven by the oscillation of the last pipeline stage. Their communication scheme is identical to the one used in S-SRSL pipelines. (iii) Delay-tolerant self-resetting stage logic (D-SRSL) Pipelines: While communication in these pipelines is local in nature in a manner similar to the one used in S-SRL pipelines, this communication is nevertheless extended in both directions. The result of this bi-directional approach is an increase in the capability of the pipeline to handle stages with random delay. Based on these pipelining techniques, a new design methodology is proposed to synthesize clockless designs. The synthesis problem consists of synthesizing an SRSL pipeline from a gate netlist with a minimum area overhead given a specified data rate. A two-phase heuristic algorithm is proposed to solve this problem. The goal of the algorithm is to pipeline a given datapath by minimizing the area occupied by inter-stage latches without violating any timing constraints. Experiments with this synthesis algorithm show that while P-SRSL pipelines can reach high throughputs in shallow pipelines, D-SRSL pipelines can achieve comparable throughputs in deeper pipelines.

As designers began to pack multi-million transistors onto a single chip, their reliance on a global clocking signal to orchestrate the operations of the chip has started to face almost insurmountable difficulties. As a result, designers started to explore clockless circuits to avoid the global clocking problem. Recently, self-resetting circuits implemented in dynamic logic families have been proposed as viable clockless alternatives. While these circuits can produce excellent performances, they display serious limitations in terms of area cost and power consumption. A middle-of-the-road alternative, which can provide a good performance and avoid the limitations seen in dynamic self-resetting circuits, would be to implement self-resetting behavior in static circuits. This alternative has been introduced recently as Self-Resetting Stage Logic and used to propose three types of clockless pipelines. Experimental studies show that these pipelines have the potential to produce high throughputs with a minimum area overhead if a suitable synthesis methodology is available. This thesis proposes a novel synthesis methodology to design and verify clockless pipelines implemented in SRSL by taking advantage of the maturity of current CAD tools. This methodology formulates the synthesis problem as a combinatorial analytical problem for which a run-time efficient exact solution is difficult to derive. Consequently, a two-phase algorithm is proposed to synthesize these pipelines from gate netlists subject to user-specified constraints. The first phase is a heuristic based on the as-soon-as-possible scheduling strategy in which each gate of the netlist is assigned to a single pipeline stage without violating the period constraint of each pipeline stage. On the other hand, the second phase consists of a heuristic, based on the Kernighan-Lin partitioning strategy, to minimize the number of nets crossing each pair of adjacent pipeline stages. The objective of this optimization is to reduce the number of latches separating pipeline stages since these latches tend to occupy large areas. Experiments conducted on a prototype of the synthesis algorithm reveal that these self-resetting stage logic pipelines can easily reach throughputs higher than 1 GHz. Furthermore, these experiments reveal that the area overhead needed to implement the self-resetting circuitry of these pipelines can be easily amortized over the area of the logic embedded in the pipeline stages. In the overall, the synthesis methods developed for SRSL produce low area overhead pipelines for wide and deep gate netlists while it tends to produce high throughput pipelines for wide and shallow gate netlists. This shows that these pipelines are mostly suitable for coarse-grain datapaths.

No description available.