Multi-chip dataflow architecture for massive scale biophysically accurate neuron simulation

Abstract

The ability to simulate brain neurons in real-time using biophysically-meaningful models is a critical pre-requisite grasping human brain behavior. By simulating neurons' behavior, it is possible, for example, to reduce the need for in-vivo experimentation, to improve artificial intelligence and to replace damaged brain parts in patients. A biophysically accurate but complex neuron model, which can be used for such applications, is the Hodgkin-Huxley (HH) model. State of the art simulators are capable of simulating, in real-time, tens of neurons, at most. The currently most advanced simulator is able to simulate 96 HH neurons in real-time. This simulator is limited by its exponential growth in communication costs. To overcome this problem, in this thesis, we propose a new system architecture, which massively increases the amount of neurons which is possible to simulate. By localizing communications, the communication cost is reduced from an exponential to a linear growth with the number of simulated neurons As a result, the proposed system allows the simulation of over 3000 to 19200 cells (depending on the connectivity scheme). To further increase the number of simulated neurons, the proposed system is designed in such a way that it is possible to implement it over multiple chips. Experimental results have shown that it is possible to use up to 8 chips and still keeping the communication costs linear with the number of simulated neurons. The systems is very flexible and allows to tune, during run-time, various parameters, including the presence of connections between neurons, eliminating (or reducing) resynthesis costs, which turn into much faster experimentation cycles. All parts of the system are generated automatically, based on the neuron connectivity scheme. A powerful simulator that incorporates latencies for on and off chip communication, as well as calculation latencies, can be used to find the right configuration for a particular task. As a result, the resulting highly adaptive and configurable system allows for biophysicallyaccurate simulation of massive amounts of cells. The undersigned hereby certify that they have read and recommend to the Faculty of Electrical Engineering, Mathematics and Computer Science for acceptance a thesis entitled "Multi-chip dataflow architecture for massive scale biophysically accurate neuron simulation" by Jaco A. Hofmann in partial fulfillment of the requirements for the degree of Master of Science. AbstractThe ability to simulate brain neurons in real-time using biophysically-meaningful models is a critical pre-requisite grasping human brain behavior. By simulating neurons' behavior, it is possible, for example, to reduce the need for in-vivo experimentation, to improve artificial intelligence and to replace damaged brain parts in patients. A biophysically accurate but complex neuron model, which can be used for such applications, is the Hodgkin-Huxley (HH) model. State of the art simulators are capable of simulating, in real-time, tens of neurons, at most. The...

Energy-efficient multipath ring network for heterogeneous clustered neuronal arrays

“…Previous effort has been made to reduce throughput and resource costs by grouping neuron simulating hardware, memory blocks and control circuitry into units referred to as clusters [11,12,13]. The reuse of hardware for computational intensive mathematical operations such as exponentiation has successfully resulted in a reduction of resource consumption without increasing overall latency.…”

A real-time hybrid neuron network for highly parallel cognitive systems

“…In previous implementations [9,17], the exponent architecture is scheduled and implemented by the Vivado HLS tools. This resulted in several large exponent architectures being implemented within the total hardware topology.…”

“…While evaluating the NN, all calculated dendrite responses are routed through a cluster controller to other cluster controllers (leaves), and the interface bridge (top node). Based on [17] a tree network is best implemented for this type of routing. Axon hillock responses are currently only used as an output, and thus do not need to be routed to the leaves, only the top node.…”

“…Each row represent where a local SimC is located in the the IO topology 17 . The row number is determined by comparing by which FIFO channel the request arrives and what job number is sent:…”

“…The network is based on work by [17], where a certain set of routers transmit information based on pre-configured routing tables. In this section the routing method and implementation for the proposed design are discussed.…”