Research On The Architecture Of Multi-CPU Neuromorphic Analog System

The development of artificial intelligence has been in full swing since Alphago defeated the top human go players in the pro go competition in 2016.However,based on the limitations of the computer architecture,it is very difficult to design human-like intelligence.The computer system of the "Von Neumann" architecture is characterized in that the memory is separated from the CPU,and the data is exchanged between the two through the data bus.When dealing with large-scale problems,the communication delay of the storage unit and the computing unit becomes a performance bottleneck,resulting in a serious wall effect.And the biological brain processes the information through the neural network,the neuron has both the calculation function and the storage function,and the integration of the operation and the storage height structure is realized,and the power consumption is very low.Therefore,the multi-CPU neuromorphic analog system is studied,the multi-CPU organization mode is designed reasonably,the multi-channel switching characteristic of the FPGA is utilized,the proper interconnection structure is used,and the physical connection of the interconnection is reduced to realize the neuromorphic analog system.In this paper,the related problems of multi-CPU neuromorphic analog system are studied.First,the multi-CPU neuromorphic analog system needs to solve the problem of how many CPUs can effectively organize the interconnection.According to the problem,the interconnection structure of the quadrilateral,the hexagon and the octagon is analyzed,and the final result selects the hexagon which has the symmetry,the expansibility and the network diameter as the interconnection structure of the computing node.Based on the characteristics of the interconnection structure of the FPGA,the CAM and the hexagon,the hardware of the neuromorphic analog system is proposed and the circuit board is developed.Secondly,the paper puts forward the routing protocol of multi-CPU neuromorphic analog system.In this paper,the fixed routing table is used to reduce the complexity of the design of the routing protocol.Third,the single-CPU single-core neuron analog software system is designed to solve the problem of the transmission of the analog neuron and the nerve impulse data packet in the thread.On the basis of a single-CPU single-core neuron simulation software system,a single-CPU multi-core neuron simulation software system is proposed to solve the problem of the transmission of the pulse data packets between the threads.In combination with that multi-CPU neuromorphic analog system main board,the single-CPU multi-core neuron simulation software system is further optimized,and the target of the multi-CPU core neuron is realized.According to the design scheme of this paper,the motherboard of neuromorphic analog system is developed to verify the feasibility of the system.The test program is designed and tested on the motherboard of neuromorphic analog system to verify the transmission performance of the system.RS,BI and CH cortical neurons are simulated and compared with the simulation provided by Eugene M.Izhikevich.Finally,the simulation experiment of Izhikevich neural network is carried out.