Since 2019, the chair of Integrated Digital Systems (IDS) participates in the “Advanced Computing Architectures” (ACA, Fig. 1) project that targets the accelerated simulation of biological neural networks. As much as science has discovered about the various electro-chemical processes in the brain, it remains still a secret of how it all adds up to create intelligence. To further advance the understanding, an essential step is the emulation of learning processes in the brain. As for a human it takes years, researchers need fast “turn-around times” to validate their assumptions in simulation. Furthermore, they would like to sweep parameters to find the best fit between the simulation results and data observed in nature.
Best-in-class HPC systems provide a simulation speed that more-or-less matches biological real-time. Within the ACA project, we target a 100x acceleration factor for biological networks of relevant complexity, i.e. covering about 10% of the human cortex down to detailed biological neuron models.
As a first step, we analyzed the challenge of communicating the spike information of the 109 neurons to its 104 ‘followers’ each. The work relies on various modelling efforts of the communication requirements and supportive hardware infrastructure [Kau20]. For benchmarking, it also requires a model of the biological communication network – the connectome (cf. Fig. 2 a sketch of such model). This simulation based exploration is supported with a physical cluster of FPGA boards (see Fig. 3 – provided by the Research Center Jülich) to emulate future system capabilities in order to calibrate the derived models.
Targeting the scientific community, it largely differs from the application oriented artificial neural networks as used for big data processing today. However, one day we foresee these efforts to converge providing artificial general intelligence to digital systems in our daily lives.
[Kau20] K. Kauth, T. Stadtmann, R. Brandhofer, V. Sobhani and T. Gemmeke, “Communication Architecture Enabling 100x Accelerated Simulation of Biological Neural Networks”, Proc. ACM/IEEE Int. Workshop on System-Level Interconnect Problems and Pathfinding (SLIPˆ2), 2020. doi: 10.1145/3414622.3431909