Contributions to the Design of Asynchronous Macromodular Systems

Plana, Luis A.

In this thesis, I advocate the use of macromodules to design and build robust and performance-competitive asynchronous systems. The contributions of the work relate to different aspects of the design of asynchronous macromodular systems. First, an architectural optimization for 4-phase systems is introduced. The goal of the optimization is to increase the performance of a system by increasing the level of concurrent activity in the sequencing of data processing stages. In particular, three new asynchronous sequencers are designed, which increase the throughput of the system. Existing asynchronous data paths do not operate correctly at this increased level of concurrency: data hazards may result. Interlock mechanisms are introduced to insure correct operation. The technique can also be regarded as a low-power optimization: The increased throughput can be traded for a significant reduction in the power consumption of the entire system. SPICE simulation results show that the new sequencers allow roughly twice the throughput of non-concurrent sequencers. The simulations also show that, after voltage scaling, energy dissipation is reduced by a factor of 2.5. Second, the use of pulses for efficient inter-module synchronization is introduced. The idea is complemented with the definition of a pulse-mode handshake protocol and the characterization of Pulse-Burst Operation (PBO), an important extension to traditional pulse-mode operation. Also, a basic set of macromodules, that efficiently implement control operations such as sequencing, selection, iteration, concurrency control, resource sharing, and arbitration is presented. Modules for interfacing pulse-mode circuits with traditional 2-phase and 4-phase circuits are also included in the set. Finally, the design of a packet switch is used to demonstrate the viability of pulse-mode macromodules to implement complex, high performance systems. The switch organization, its asynchronous operation, and the low control overhead introduced by pulse-mode macromodules result in a design that can handle 2.4 times the target throughput of 155 Mbits/Sec. Also, the switch is characterized by very low input-to-output latency. These results suggest that pulse-mode macromodules can keep control overhead low without introducing complex, unsafe timing considerations, two necessary conditions to achieve robust, performance-competitive systems.



More About This Work

Academic Units
Computer Science
Department of Computer Science, Columbia University
Columbia University Computer Science Technical Reports, CUCS-001-99
Published Here
April 21, 2011