Asynchronous VLSI

Event-driven computing

In the field of neuromorphic computing where millions of neurons operate with extremely sparse, data-dependent activity. A data-structure is often used to record which neuron produced an output (”fired”) and then use this information in the next iteration of processing. Common approaches to manage such events either implements some fixed-length bit-vector, which is efficient when the event set is densely populated, or they implement some type of dynamic queue, which is efficient for extremely sparse event set. We proposed a new solution for managing moderately sparse event set and designed two implementation with different trade-offs.

Event-driven communication

Silicon neuromorphic systems have a large, distributed array of computation units (called neurons) that operate with extremely sparse, data-dependent activity at relatively slow timescales. Communicating events between these computation units with direct, point-to-point wiring is difficult for a large-scale system. The standard approach to this problem is to leverage the speed of modern CMOS and use time-multiplexed wires for communication. Address-event representation (AER) is an example of such event-driven encoding and communication protocol which was originally proposed to communicate the location and timing information of sparse neural events between neuromorphic chips. A variety of schemes have been proposed in the literature for designing an AER-based communication interface based on the event or activity pattern. 

We designed two new encoding schemes which offer a significant improvement in latency, throughput, and power compared to existing approaches.

For situations where the event (activity) pattern is unknown, HTR  handles event with the help of a hierarchical ring structure.

When the event (activity) pattern is known, FP-AER configures into a topology that works best for the given activity pattern.

Event-driven sensing

(details coming soon...)