Self-Timed Router for a Native TrueNorth
Abstract
The human brain is capable of highly energy-efficient information processing. Estimates are that the human brain uses about 20 Watts of power and occupies about 1.2L of volume. Within this space and power budget we routinely perform visual processing tasks like real-time object detection, auditory processing tasks like speaker recognition, being able to pick out individual voices in a group setting, etc.—tasks that are highly challenging from a computational perspective on conventional computing substrates like general-purpose microprocessors. Inspired by this, neuromorphic electronics attempts to capture the essence of neural information processing by mimicking biological neurons, synapses, and their connectivity in a silicon implementation. In a six year collaborative research effort with IBM Research, Cornell has developed the programmable TrueNorth architecture for neuromorphic computing [1]. TrueNorth represents a major milestone in the development of neuromorphic computing architectures. It is a single-chip million neuron, 256 million synapse chip with a power budget of 70 mW—significantly lower than other neuromorphic systems, including the proposed work for the European Human Brain Project. TrueNorth takes the neuromorphic paradigm to its limit—all primary inputs and outputs are encoded as spike trains and communicated via a self-timed routing interface, as are all inter-neuron communications on-chip and between TrueNorth chips. Almost all conventional computers are implemented using the clocked design paradigm. Hence, augmenting a conventional computing platform with the TrueNorth capability requires interfacing to TrueNorth’s self-timed interface. Using conventional clocked logic to implement this interface can lead to overheads, because sequencing in clocked logic is typically performed at the granularity of individual clock cycles—many more gates than the sequencing that can be implemented with self-timed logic. The objective of the proposed task is to work collaboratively with AFRL to develop a custom interface integrated circuit to TrueNorth—TNIC (for TrueNorth Network Interface Chip)—that converts the TrueNorth interface into a conventional, synchronous interface that is familiar to users of the TrueNorth chip.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 11, 2016
- Source ID
- FA87501510173
Entities
People
- Rajit Manohar
Organizations
- Cornell University
- Rome Laboratory
- United States Air Force