Emulation of Bursting Neurons in Neuromorphic Hardware based on Phase-Change Materials (PDF)

In the history of computing hardware,Moore's law, named after Intel co-founder Gordon E. Moore, describes a long-termtrend, whereby the number of transistors that can be placed inexpensively on an integrated circuit doubles approximately every two years [1]. Because the number of transistors is crucial for computing performance, significant performance gains could be achieved simply through complementary metal-oxide-semiconductor (CMOS) transistor downscaling. AlthoughMoore's law, which was mentioned for the first time in 1965, turned out to persist for almost five decades, the nano era poses significant problems to the concept of downscaling [2]. Upon approaching the size of atoms, quantumeffects, such as quantum tunneling, pose fundamental barriers to the trend. Furthermore, the conventional computing paradigm based on the Von-Neumann architecture and binary logic becomes increasingly inefficient considering the growing complexity of todays computational tasks. Hence, new computational paradigms and alternative information processing architectures must be explored to extend the capabilities of future information technology beyond digital logic. A fantastic example for such an alternative information processing architecture is the human brain. The brain provides superior computational features such as ultrahigh density of processing units, low energy consumption per computational event, ultrahigh parallelism in computational execution, extremely flexible plasticity of connections between processing units and fault-tolerant computing provided by a huge number of computational entities. Compared to today's programmable computers, biological systems are six to nine orders of magnitude more efficient in complex environments [3]. For instance: simulating five seconds of brain activity takes IBM's state-of-the-art supercomputer Blue Gene a hundred times as long, i.e. 500 s, during which it consumes 1.4 MWof power, whereas the power dissipation in the human central nervous system is of the order of 10W[4, 5]. Thus, it is not only extremely interesting but in terms of computational progress also highly desirable to understand how information is processed in the human brain. The conceptual idea developed within the framework of this thesis tries to contribute to this intention. In contrast to most recent research dealing with the simulation and emulation of specific connections between nerve cells [5-12], the work of this thesis focuses on investigating, on [...]