Synchronous Elastic Circuits

Synchronous elastic circuits (also known as latency-insensitive and latency-tolerant) behave independently of the latencies of computations and communication channels. For example, the three sequences are an acceptable behavior of an elastic adder with input channels X, Y and output channel Z, where the absence of transfer on a particular channel at a given cycle is indicated by *. Indeed, the associated transfer subsequences (obtained by deleting the * 's) make up a behavior of an ordinary (non-elastic) adder: Current interest in elasticity is motivated by the difficulties with timing and communication in large synchronous designs in nanoscale technologies. The time discretization imposed by synchronicity forces to take early decisions that often complicate changes at the latest stages of the design or efficient design scaling. In modern technologies, calculating the number of cycles required to transmit data from a sender to a receiver is a problem that often cannot be solved until the final layout has been generated. Elastic circuits promise novel methods for microarchitectural design that can use variable latency components and tolerate static and dynamic changes in communication latencies, while still employing standard synchronous design tools and methods. We will first present a simple elastic protocol, called SELF (Synchronous Elastic Flow) and describes methods for an efficient implementation of elastic systems and for the conversion of regular synchronous designs into an elastic form. Every elastic circuit E implements the behavior of an associated standard (non-elastic) circuit C, as in the adder example above. For each wire X of C, there are three in E: the data wire D X , and the single-bit control wires V X and S X (valid and stop). This triple of wires is a channel of E. A transfer along the channel occurs when V X = 1 and S X = 0, thus requiring cooperation of the producer and the consumer. [CKG06] provides more details on the implementation of SELF. We will next review theoretical foundations of SELF. Our main result states that (under favorable circumstances) " the network of elasticizations is an elasti