A Low-Latency Communication Design for Brain Simulations

Brain simulation, as one of the latest advances in artificial intelligence, facilitates better understanding about how information is represented and processed brain. The extreme complexity human brain makes simulations only feasible on high-performance computing platforms. Supercomputers with a large number interconnected graphical processing units (GPUs) are currently employed for supporting simulations. Therefore, high-throughput low-latency inter-GPU communications super-computers play crucial role meeting performance requirements simulation highly time-sensitive application. In this article, we first provide an overview current parallelizing technologies using multi-GPU architectures. Then analyze challenges to summarize guidelines communication design address such challenges. Furthermore, propose partitioning algorithm two-level routing method achieve efficient architecture simulation. We report experiment results obtained supercomputer 2000 GPUs simulating model 10 billion neurons (digital twin brain, DTB) show that our approach can significantly improve performance. also discuss open issues identify some research directions