Survey of Advancements in Non Von NeumannArchitecturesAmeed FarabiNational University of Computerand Emerging SciencesEmail: [email protected]—Progress made in architecture design in pursuit ofmassively parallel Non von Nuemann based machines catering themodern demands of data rich and data centric applications doesnot have a single point of focus and performance improvementsare sought after using many both hardware based and natureinspired solutions.
The advancements have been made mainly inneuromorphic information systems design, probabilistic process-ing and hardware circuits capable of processing and storing dataat the same time.Index Terms—non von architectures, massively parallel archi-tectures and neuromorphic computing.I. I NTRODUCTIONIn this research we explore the current state of the art inNon Von architecture design also called beyond Non VonComputing. The developments in this area that we havecovered for this study can be narrowed down to following:- Neuromorhic information processing systems also knownas the brain inspired designs. – Hardware circuit designs tobring processing and memory at the same place. – Parallelautomata processing and associative processing. – Probabilisticmachine designs.
Most Non Von designs and the research doneto achieve massive levels of parallelism is aimed at solvingspecific challenges presented by applications that require suchhuge levels of parallelism like cognitive computing applica-tions, machine learning applications and artificially intelligentapplications.II. B ACKGROUNDVon Neumann bottleneck is the biggest problem limitingthe operational capacity and speed of systems designed basedVon Neumann architecture.
As shown in the figure CPU andmemory are connected by a bus therefore the throughput ofCPU is greatly affected by the bus capacity and speed.Despite this drawback Von Neumann architecture waswidely adopted and the overall performance of computing de-vices was improved by other means such as pipelining, cache,task level parallelism and many more. Also the applicationsrunning on most machines did not demand great degrees ofparallelism unlike most modern applications which are datacentric.
The most important aspect that masked this drawbackwas the constant increase in number of transistors in chipswhich has also now stagnated and is set to decline in future,this phenomenon is better known as death of Moors law13.In general the different developments in overall architecturedesign since the adoption of Von Neumann did not focus toalter the shared bus arrangement but to compensate it fromreaping benefits of other developments as mentioned above.But the modern applications demand a fundamental change tothis arrangement and research to bring forth newer and betterNon Von Neumann based architectures has been on the rise.III. N EUROMORPHIC I NFORMATION S YSTEMS :Neuromorphic information processing systems are devel-oped by applying design aspects of human nervous sys-temMIP.
These are in form of hardware circuits and arerequired to posses following qualities that are found in ahuman brain: parallel information processing, fault tolerance,low power requirements, adaptive and event driven. Whatdifferentiates a Neuromorphic information system from aconventional system is that conventional systems have memoryand computation components separated by a communicationbus while for a Neuromorphic information system the memoryand processing unit are co-located. There are two majorschools for design of these systems: computationalism andconnectionalism. The former advocates that the behavior re-lated information should be stored at intersection points ornodes (neurons) while the later advocates that the channel ofcommunication should contain the behavior related informa-tion.A. True North:A brain inspired processor targeting the kind performanceattributed to the human brain like communication betweenmillions of neurons and responding to parallel events inte-grated by very tightly connected event based communicationinfrastructure. To achieve this kind of nature like brillianceon a silicon chip the 4096 neurosynaptic cores are integratedand it contains one million digital neurons.
The behaviorof neurons and their connectivity details are provided andaccordingly the processor is programmed. The brain only takesonly 20W to operate and perform all its jobs, TrueNorth alsotargets this kind of low operational power and compared toclassical Von Neumann based machines its tested to utilizemuch less power to execute same amount of processes. Theapplication of this processor are mainly in the area of imagerecognition simulating the important function of its benchmarkthe brain which has the ability to recognize images and basedon it allow other parts of body controlled by it to behaveaccordingly.
IV. B AYESIAN MACHINESIntelligent systems however detailed and well tested stillface the challenge of their systems response to uncertainty. Atthe same time living beings regularly deal with uncertaintyand their experiences drive their future behaviors to similarevents. This ability when modeled into an intelligent systemmay enable better behavior to uncertainty. This would requirehardware circuits to be developed on the lines of informationgained from Bayesian inference. The basic idea here is thatBayesian inferences are used to describe the circuit design andsuch a circuit would be better suited to run in an intelligentsystem dealing with uncertain inputs.A bayesian inference model has been developed as partof BAMBI project and its goals are described as follows: 1.Uncertainty of inputs should be considered.
2. Should be faulttolerant. 3.
Should require low power.V. H ARDWARE A DVANCES E NABLING D ESIGN OF N ONV ON M ACHINES :A. MTJ:In their research Friedman et al. propose a logic family inwhich logic gates can be cascaded directly. This is achieved byinserting magnetic tunnel junctions (MTJ) in circuit structure.Logical operations can be performed and also data is encodedin long term storage.
This circuit design allows bringingcomputation with memory which is important feature for nonvon machine design.B. MAGIC:Memristors have variable resistance based on previousvoltages applied. Usually they are used as memory but theyprovide great potential for logic circuits. A state is representedas resistance considering high and low voltages. The input andoutput states are included using separate memristors.
Thesevariable voltages used as states are suitable to save informationand context in digital neurons.C. Advance Materials:Phase change materials have an important previously unexplored property called accumulation which enable the pos-sibilities of performing arithmetic computations. This proposesopportunity for designing Non von systems as at the time offiring of events there is not only information about the memorybut also the neuron can respond to other events by performingcomputations.D.
On chip Learning:The ability to parallel read and write across large numberof non volatile memory units can increase parallelism and issuitable for neuromorphic computing models. Forward propa-gation is a concept presented through which firing of neuronevents can be computed keeping account of context of previousactivations.VI. R EFERENCES