Large-scale fan-out and fan-in can be achieved in single flux quantum architectures but they require a commensurate number of additional JJ elements. Another critical departure of the human brain from typical modern computer architectures is its extremely high connectivity.
This energy advantage holds true even after taking into account the cooling requirements to operate at 4.2 K, which can be less than 1000 W to cool 1 W at 4.2 K. In addition, the energy consumption of spiking JJs can readily be below 10 −18 J, also giving them an advantage compared to the energy needed to produce a spike in the human brain, which typically requires on the order of 10 −14 J. Further, one particularly compelling trait of neuromorphic JJ circuits is the potential to operate at tens or hundreds of gigahertz achieving many orders of magnitude in speed over the human brain, which typically operates below 1 kHz. However, many potential applications do not require such a large junction count and therefore seem tractable with the current state of JJ fabrication technology. These numbers exceed the largest JJ circuits, which are on the order of 10 6. The number of neurons in the brain is on the order of 10 11 and the number of synapses is on the order of 10 14. We discuss architectures that explicitly set the weights of these artificial synapses as well as those that make use of hidden-layer dynamics whose internal weights are less important than their overall topology. The memory and plasticity mechanisms in the brain, which are implemented with modulations of the synaptic strength between neurons, can also be implemented with a simple two-JJ circuit or a novel magnetic JJ, among other options. The near lossless transmission of voltage spikes that occur in the human brain can be easily implemented with active or passive superconducting transmission lines. More biologically realistic spiking behavior including following the basic ion flow dynamics of Na and Ca within a neuron can be implemented with a two JJ circuit. Examples of these primitives include the spiking nature of neurons in the human brain, which can be mimicked with a single JJ. Josephson junctions (JJs) are naturally neuromorphic hardware devices that can implement several biological primitives at the device level. While such advances in digital superconducting systems are very important, in this review we survey the analog and mixed analog–digital use of biologically inspired computing design from the device up, and its potential to build an efficient neuromorphic computational stack. These systems compare favorably in both energy and speed compared with purpose built CMOS hardware designed for the same task.
#Magnetic supermind software#
It should be noted that recent work on digital superconducting architectures has shown that significant improvements can be made in the design of superconducting systems that can train and run software based neural networks. However, these algorithms are still predominantly trained and run on hardware that is optimized for digital logic with its associated memory bottleneck and high power consumption.
Because of the success of these algorithms, the computational need for training them has been doubling every four months, nearly five times the exponential growth rate offered by Moore's Law operating at its peak.
The potential power of biologically inspired computing is evidenced in the growing prevalence of artificial intelligence or machine learning algorithms. Neuromorphic computing seeks to take advantage of several of these characteristics by implementing biologically inspired design directly in hardware. The brain is a highly parallel computing platform that exhibits tremendous adaptability and fault tolerance while simultaneously being extremely energy efficient.
The human brain offers some insight into a method of computation that differs from typical digital logic.
0 kommentar(er)
