What do complex social infrastructure systems and Artificial Intelligence systems have in common? They both impose enormous and expanding computational loads. Given enough time and resources these computational loads might be successfully processed using conventional computing architectures. However, as the complexity and volume of these computational loads increases, at some point it will become economically and technically unfeasible to continue with conventional computing.
What if you could move specific computational problem solving loads to a chip that was designed to handle these complex problems in a much more efficient manner and at a fraction of the foot print and cost?
Enter Hitachi’s Ising chip. Rather than executing the problem solving procedures sequentially as with conventional compute architectures, Hitachi is proposing a different concept called “natural computing” as applied to an Ising model.
At this point we need to understand the following key terms:
1. Combinatorial optimization
Combinatorial optimization refers to a class of mathematical problems where a solution must be found that maximizes (or minimizes) a performance index under given conditions. A characteristic of combinatorial optimization problems is that the number of candidate solutions increases explosively as the number of parameters increases. A classic example of this kind of problem is the “traveling salesperson problem”, in this problem the salesperson has a list of cities that must be visited, and the problem to be solved is “what is the most efficient travel route I can take through all cities to minimize my travel time?”. What is being sought is the best of many possible answers. As more cities are added to the list the number of possible answers grows dramatically. The social infrastructure and AI technologies can generate very large combinatorial optimization problems.
2. Natural computing
Quoting Springer Intl. Publishing, publisher of Natural Computing Journal: http://link.springer.com/journal/11047
"Natural Computing refers to computational processes observed in nature, and human-designed computing inspired by nature. When complex natural phenomena are analyzed in terms of computational processes, our understanding of both nature and the essence of computation is enhanced. Characteristic for human-designed computing inspired by nature is the metaphorical use of concepts, principles and mechanisms underlying natural systems. Natural computing includes evolutionary algorithms, neural networks, molecular computing and quantum computing."
3. The Ising model
The Ising model (named after Ernst Ising) is a mathematical model concerned with the physics of phase transitions, which occur when a small change in a parameter causes a large-scale, qualitative change in the state of a system. The properties of a magnetic material are determined by magnetic spins, which can be oriented up or down. An Ising model is expressed in terms of the individual spin states, the interaction coefficients that represent the strength of the intersections between different pairs of spin states, and the external magnetic coefficients that represent the strength of the external magnetic field.
What has Hitachi Research done?
Instead of heating up the data center running countless permutations of variable combinations looking for the optimal combination of those variables using conventional computing, Hitachi is proposing to use a method for natural computing and use a natural phenomenon to model the problem to be solved (mapping) and take advantage of the convergence inherent in this natural phenomenon to converge on the solution to the problem.
Hitachi proposes replicating the Ising model using a Complementary Metal Oxide Semiconductor (CMOS) circuit. Hitachi maps the combinatorial problem into the Ising model in the CMOS circuit in such a way that its performance index corresponds to the model’s energy, the Ising model is allowed to converge so that the spin state adopts the minimum energy configuration. This minimum energy configuration is equivalent to obtaining the optimal combination of parameters that minimizes the performance index of the original optimization problem.
Early prototypes have proven themselves to be effective and as much as 1800x more energy efficient when compared to competing Ising computer technologies. Hitachi continues to innovate in this area to address the growing computational demands of AI and social infrastructure systems. If you would like to learn more about this Hitachi Research activity, I recommend the following articles published in the Hitachi Review magazine: