The new era of cognitive computing brings forth the grand challenge of developing systems capable of processing massive amounts of noisy multisensory data. This type of intelligent computing poses a set of constraints, including real-time operation, low-power consumption and scalability, which require a radical departure from conventional system design. Brain-inspired architectures offer tremendous promise in this area. To this end, we developed TrueNorth, a 65 mW real-time neurosynaptic processor that implements a non-von Neumann, low-power, highly-parallel, scalable, and defect-tolerant architecture. With 4096 neurosynaptic cores, the TrueNorth chip contains 1 million digital neurons and 256 million synapses tightly interconnected by an event-driven routing infrastructure. The fully digital 5.4 billion transistor implementation leverages existing CMOS scaling trends, while ensuring one-to-one correspondence between hardware and software. With such aggressive design metrics and the TrueNorth architecture breaking path with prevailing architectures, it is clear that conventional computer-aided design (CAD) tools could not be used for the design. As a result, we developed a novel design methodology that includes mixed asynchronous–synchronous circuits and a complete tool flow for building an event-driven, low-power neurosynaptic chip. The TrueNorth chip is fully configurable in terms of connectivity and neural parameters to allow custom configurations for a wide range of cognitive and sensory perception applications. To reduce the system’s communication energy, we have adapted existing application-agnostic very large-scale integration CAD placement tools for mapping logical neural networks to the physical neurosynaptic core locations on the TrueNorth chips. With that, we have successfully demonstrated the use of TrueNorth-based systems in multiple applications, including visual object recognition, with higher performance and orders of magnitude lower power consumption than the same algorithms run on von Neumann architectures. The TrueNorth chip and its tool flow serve as building blocks for future cognitive systems, and give designers an opportunity to develop novel brain-inspired architectures and systems based on the knowledge obtained from this paper.

TrueNorth: Design and Tool Flow of a 65 mW 1 Million Neuron Programmable Neurosynaptic Chip

In this paper, we describe the design of Neurogrid, a neuromorphic system for simulating large-scale neural models in real time. Neuromorphic systems realize the function of biological neural systems by emulating their structure. Designers of such systems face three major design choices: 1) whether to emulate the four neural elements-axonal arbor, synapse, dendritic tree, and soma-with dedicated or shared electronic circuits; 2) whether to implement these electronic circuits in an analog or digital manner; and 3) whether to interconnect arrays of these silicon neurons with a mesh or a tree network. The choices we made were: 1) we emulated all neural elements except the soma with shared electronic circuits; this choice maximized the number of synaptic connections; 2) we realized all electronic circuits except those for axonal arbors in an analog manner; this choice maximized energy efficiency; and 3) we interconnected neural arrays in a tree network; this choice maximized throughput. These three choices made it possible to simulate a million neurons with billions of synaptic connections in real time-for the first time-using 16 Neurocores integrated on a board that consumes three watts.

/pdf/neurogrid-a-mixed-analog-digital-multichip-system-for-large-494ipl0j1g.pdf

Neurogrid: A Mixed-Analog-Digital Multichip System for Large-Scale Neural Simulations

Principles of Asynchronous Circuit Design - A Systems Perspective addresses the need for an introductory text on asynchronous circuit design. Part I is an 8-chapter tutorial which addresses the most important issues for the beginner, including how to think about asynchronous systems. Part II is a 4-chapter introduction to Balsa, a freely-available synthesis system for asynchronous circuits which will enable the reader to get hands-on experience of designing high-level asynchronous systems. Part III offers a number of examples of state-of-the-art asynchronous systems to illustrate what can be built using asynchronous techniques. The examples range from a complete commercial smart card chip to complex microprocessors. The objective in writing this book has been to enable industrial designers with a background in conventional (clocked) design to be able to understand asynchronous design sufficiently to assess what it has to offer and whether it might be advantageous in their next design task.

/pdf/principles-of-asynchronous-circuit-design-a-systems-58fhcojs1t.pdf

Principles of Asynchronous Circuit Design: A Systems Perspective

This work observes that a large fraction of the computations performed by Deep Neural Networks (DNNs) are intrinsically ineffectual as they involve a multiplication where one of the inputs is zero. This observation motivates Cnvlutin (CNV), a value-based approach to hardware acceleration that eliminates most of these ineffectual operations, improving performance and energy over a state-of-the-art accelerator with no accuracy loss. CNV uses hierarchical data-parallel units, allowing groups of lanes to proceed mostly independently enabling them to skip over the ineffectual computations. A co-designed data storage format encodes the computation elimination decisions taking them off the critical path while avoiding control divergence in the data parallel units. Combined, the units and the data storage format result in a data-parallel architecture that maintains wide, aligned accesses to its memory hierarchy and that keeps its data lanes busy. By loosening the ineffectual computation identification criterion, CNV enables further performance and energy efficiency improvements, and more so if a loss in accuracy is acceptable. Experimental measurements over a set of state-of-the-art DNNs for image classification show that CNV improves performance over a state-of-the-art accelerator from 1.24× to 1.55× and by 1.37× on average without any loss in accuracy by removing zero-valued operand multiplications alone. While CNV incurs an area overhead of 4.49%, it improves overall EDP (Energy Delay Product) and ED2P (Energy Delay Squared Product) on average by 1.47× and 2.01×, respectively. The average performance improvements increase to 1.52× without any loss in accuracy with a broader ineffectual identification policy. Further improvements are demonstrated with a loss in accuracy.

Cnvlutin: ineffectual-neuron-free deep neural network computing

This paper discusses connectivity between neuromorphic chips, which use the timing of fixed-height fixed-width pulses to encode information. Address-events (log/sub 2/(N)-bit packets that uniquely identify one of N neurons) are used to transmit these pulses in real time on a random-access time-multiplexed communication channel. Activity is assumed to consist of neuronal ensembles-spikes clustered in space and in time. This paper quantifies tradeoffs faced in allocating bandwidth, granting access, and queuing, as well as throughput requirements, and concludes that an arbitered channel design is the best choice. The arbitered channel is implemented with a formal design methodology for asynchronous digital VLSI CMOS systems, after introducing the reader to this top-down synthesis technique. Following the evolution of three generations of designs, it is shown how the overhead of arbitrating, and encoding and decoding, can be reduced in area (from N to /spl radic/N) by organizing neurons into rows and columns, and reduced in time (from log/sub 2/(N) to 2) by exploiting locality in the arbiter tree and in the row-column architecture, and clustered activity. Throughput is boosted by pipelining and by reading spikes in parallel. Simple techniques that reduce crosstalk in these mixed analog-digital systems are described.

/pdf/point-to-point-connectivity-between-neuromorphic-chips-using-2x6lccyaeo.pdf

Point-to-point connectivity between neuromorphic chips using address events

The design of an asynchronous clone of a MIPS R3000 microprocessor is presented. In 0.6 /spl mu/m CMOS, we expect performance close to 280 MIPS, for a power consumption of 7 W. The paper describes the structure of a high-performance asynchronous pipeline, in particular precise exceptions, pipelined caches, arithmetic, and registers, and the circuit techniques developed to achieve high throughput.

The design of an asynchronous MIPS R3000 microprocessor

SoC design will require asynchronous techniques as the large parameter variations across the chip will make it impossible to control delays in clock networks and other global signals efficiently. Initially, SoCs will be globally asynchronous and locally synchronous (GALS). But the complexity of the numerous asynchronous/synchronous interfaces required in a GALS will eventually lead to entirely asynchronous solutions. This paper introduces the main design principles, methods, and building blocks for asynchronous VLSI systems, with an emphasis on communication and synchronization. Asynchronous circuits with the only delay assumption of isochronic forks are called quasi-delay-insensitive (QDI). QDI is used in the paper as the basis for asynchronous logic. The paper discusses asynchronous handshake protocols for communication and the notion of validity/neutrality tests, and completion tree. Basic building blocks for sequencing, storage, function evaluation, and buses are described, and two alternative methods for the implementation of an arbitrary computation are explained. Issues of arbitration, and synchronization play an important role in complex distributed systems and especially in GALS. The two main asynchronous/synchronous interfaces needed in GALS-one based on synchronizer, the other on stoppable clock-are described and analyzed.

/pdf/asynchronous-techniques-for-system-on-chip-design-4t29mjby2z.pdf

Asynchronous Techniques for System-on-Chip Design

We investigate an efficiency metric for VLSI computation that includes energy. E, and time, t, in the form Et2. We apply the metric to CMOS circuits operating outside velocity saturation when energy and delay can be exchanged by adjusting the supply voltage; we prove that under these assumptions, optimal Et2 implies optimal energy and delay. We give experimental and simulation evidences of the range and limits of the assumptions. We derive several results about sequential, parallel, and pipelined computations optimized for Et2, including a result about the optimal length of a pipeline.We discuss transistor sizing for optimal Et2 and show that, for fixed, nonzero execution rates, the optimum is achieved when the sum of the transistor-gate capacitances is twice the sum of the parasitic capacitances--not for minimum transistor sizes. We derive an approximation for Etn (for arbitrary n) of an optimally sized system that can be computed without actually sizing the transistors; we show that this approximation is accurate. We prove that when multiple, adjustable supply voltages are allowed, the optimal Et2 for the sequential composition of components is achieved when the supply voltages are adjusted so that the components consume equal power. Finally, we give rules for computing the Et2 of the sequential and parallel compositions of systems, when the Et2 of the components are known.

/pdf/et-2-a-metric-for-time-and-energy-efficiency-of-computation-5jiyess6aj.pdf

ET 2 : a metric for time and energy efficiency of computation

We describe the Lutonium, an asynchronous 8051 microcontroller designed for low Et/sup 2/. In 0.18 /spl mu/m CMOS, at nominal 1.8 V, we expect a performance of 0.5 nJ per instruction at 200 MIPS. At 0.5 V, we expect 4 MIPS and 40 pJ/instruction, corresponding to 25,000 MIPS/Watt. We describe the structure of a fine-grain pipeline optimized for Et/sup 2/ efficiency, some of the peripherals implementation, and the advantages of an asynchronous implementation of a deep-sleep mechanism.

The Lutonium: a sub-nanojoule asynchronous 8051 microcontroller

We trace the evolution of Caltech asynchronous processors from a simple proof of concept, to a high-performance MIPS-like processor using a different buffer circuit for better performance, to the latest 8051 clone targeting low-energy operation. We describe the control aspects of the evolving circuit styles. We describe these three generations of asynchronous microprocessors (Caltech asynchronous processors, MiniMIPS and Lutonium) and the corresponding circuit families and design methods. The asynchronous circuits we use are called quasidelay-insensitive (QDI) circuits. A QDI circuit involves no assumption about, or knowledge of, delays in operators and wires, except for isochronic forks, which the designer assumes have similar delays on the different branches. QDI circuits are the most conservative asynchronous circuits in terms of delays.

Mika Nyström

Papers

The design of an asynchronous MIPS R3000 microprocessor

Asynchronous Techniques for System-on-Chip Design

ET 2 : a metric for time and energy efficiency of computation

The Lutonium: a sub-nanojoule asynchronous 8051 microcontroller

Three generations of asynchronous microprocessors