Another fascinating property of the TrueNorth chip is its scalability. When it comes to the brain, adding more neurons and synapses to the neural tissue would seamlessly result in more computation power, where adding more cores to the processor in computers, as an analog to this matter, requires some extra work for synchronization and proper resource sharing. TrueNorth achieves scalability by tiling cores within a chip, using peripheral circuits to tile chips, and locally generating core execution signals (eliminating the global clock skew challenge). This structure would also allow circuit level (memory) redundancies, as well as disabling and routing around faulty cores at the architecture-level to provide robustness to manufacturing defects.
Applications
TrueNorth comes along with its own programming language called Corelet, an object-oriented compositional language for developing efficient, modular, scalable, reusable, and collaborative neuro-synaptic software \cite{Amir_2013}. Corelet provides a logical representation of the neuro-synaptic cores that allows for the optimization of circuits for bandwidth and active-power minimization. To further investigate the capabilities of the TrueNorth chip, several streaming video applications were deployed with astonishing results. An example of these applications is a grid classifier that attempts to detect and classify all objects in an entire image simultaneously. This application required over 62,000 cores spread over 16 separate chips but only required 2.5 W of power \cite{Cassidy_2014}! Figure \ref{326775} illustrates some of the results of this application which attained a precision of 0.82 and 0.71 recall.