The Amulet chips: Architectural development for asynchronous microprocessors

During the 1990s a series of asynchronous microprocessors based on the ARM architecture was developed at the University of Manchester. The objective was to demonstrate that it was feasible to implement a commercial architecture with asynchronous logic and that certain advantages could be gained from a self-timed processor. By carrying these designs through to silicon it was demonstrated that processors, caches and whole systems-on-chip could be built without clocks and could perform competitively with 'conventional', synchronous systems. These processors, named Amulet1-3, exhibited some useful characteristics, particularly their ability to stop and start almost instantaneously with consequent energy savings in embedded applications and their low-energy and wide spectrum electromagnetic emission useful in wireless applications. As might be expected they also adapt automatically to supply voltage variation so that their speed and energy demands may be regulated simply by altering their power supply. There were disadvantages too, particularly in the design effort which was not well supported by conventional tool flows. Thus, at that time, the advantages were not great enough for any widespread commercial acceptance. This paper summarises the architectural features which were developed during the Amulet series of microprocessors and systems with emphasis on mechanisms which are non-specific to the ARM architecture. These include register forwarding and cache-line fetching which typically rely on non-local synchronisation provided by a clock signal. In addition deadlocks, an ever-present danger in asynchronous systems, are discussed and means of preempting them presented.