Interconnection system for the spiNNaker biologically inspired multi-computer

SpiNNaker is a large-scale multi-core computing engine designed to model heavily distributed fine-grain problems. The machine is constructed hierarchically: 1 monitor and 16 worker processors form a single node of a toroidal compute `surface´. The six high-speed bi-directional links of each node are used for triangular edge connections that provide alternative routes around problematic links. The system itself is scalable from one node up to 2¹⁶ resulting in a maximum of 2²⁰ worker processors. SpiNNaker is an isotropic homogeneous network of processors that deliberately includes no central overseer. A consequence of this isotropy is an absence of perimeter and hence no natural position for peripheral I/O connections. This study describes the practical techniques and details employed in two components of the system: (a) SpiNNlink is the proposed board-to-board interconnection system that multiplexes 48 separate 250 Mbps SpiNNaker links through six off-board connections without compromising the overall system bisection bandwidth, forms an isotropic meta-network on top of SpiNNaker without requiring any cooperation from system software, and remains transparent to the SpiNNaker network; and (b) SpiNNterceptor is the proposed peripheral I/O subsystem developed as a layer on top of SpiNNlink that provides over 18 Gbps of minimally disruptive communication between SpiNNaker applications and externally connected equipment.