Communication architecture enabling 100x accelerated simulation of biological neural networks

Abstract

To further develop the understanding of cognitive processes in the human cortex, neuroscientists seek to simulate relevant biological neural networks in the order of 109 neurons with natural densities of 104 synapses per neuron. To observe long-term effects of learning, a speed-up of at least 100x with respect to biological real-time is required while preserving deterministic results and a high temporal resolution of 0.1 ms. In this paper, we translate these objectives to requirements for the communication architecture of a large-scale neuroscience simulator. These requirements are based on a connectivity model that includes gray and white matter as well as clustered connections and represents essential communication requirements of biological neural networks. In analytical and numerical analysis, existing platforms fall short of meeting all requirements simultaneously even assuming modern high-speed transceivers. This paper presents a balanced multi-hop communication architecture that cuts latency and achieves high bandwidth efficiency. Extrapolating from physical measurements of link performance, our work brings the challenging communication requirements within reach of next generation large-scale neuroscience simulation platforms.