SpiNNaker-A million core ARM-powered Neural HPC

Опубликовано 12 августа 2016, 2:39

SpiNNaker is a specialised massively-parallel computing architecture developed primarily to model large-scale spiking neural networks in real-time, but is fully programmable permitting its adaptation to many other parallel processing tasks. Its architecture is inspired by that of the biological brain, offering high-levels of computational parallelism, connectivity and redundancy in an energy efficient environment. The main building block, the SpiNNaker node, is based around a Multi-Processor System-on-Chip incorporating 18 ARM processor subsystems, packaged together with a 128Mbyte SDRAM.A SpiNNaker system may scale to a maximum of 65,536 nodes, giving rise a systems which may exceed 1 million cores, but use less than 100 kW while in-flight, modeling ~1 of the neuron count of the human brain.In the talk Cameron will discuss the hardware philosophy, architecture and implementation of SpiNNaker systems, together with the particular challenges in scaling it towards full-sized deployments. It will also address the significant software and management issues, offering approaches that have been developed to address these areas as the systems increase in size towards a million core machine.