Atomic computing - a different perspective on massively parallel problems.

As the size of parallel computing systems inexorably increases, the proportion of resource consumption (design effort, operating power, communication and calculation latency) absorbed by 'non-computing' tasks (communication and housekeeping) increases disproportionally. The SpiNNaker (Spiking neural net architecture) engine [1,2] sidesteps many of these issues with a novel architectural model: it is an isotropic 'mesh' of (ARM9) cores, connected via a hardware communication network. The topology allows uniform scalability up to a hard limit of just over a million cores, and the communications network - hardware handling packets of 72 bits - achieves a bisection bandwidth of 5 billion packets/s. The state of the machine is maintained in over 8TB of 32-bit memory, physically distributed throughout the system. There is no central processing 'overseer' or synchronised clock. This paper discusses opportunities and challenges in applying the SpiNNaker architecture, within neural simulation and beyond.