Abstract: Neuromorphic computing systems that mimic biological neural structures are being pursued as one of the likely non-Von Neumann systems that offer extremely high performance relative to power consumption and size. One of the technical challenges that need to be addressed is building an accurate model to understand how the brain works. However, existing neuroscience software have limited capability to simulate multiple regions of the brain because of memory consumption to store large networks and scalability issues.

We have been developing a scalable brain simulator using PETSc DMNetwork to scale the network size and physics needed to obtain the simulation detail for capturing the ability to learn. It provides capabilities for setting up the most common network connectivity in mammalian brains in parallel, spiking neuron and synapse models and solves the coupled ODEs of the models using PETSc Time-Integrator. In this talk, the scalability, memory consumption, power usage of its algorithms on CPU and GPUs, and future directions will be discussed.