Research Software Engineers Andrew Rowley and Donal Fellows have been seconded from Research IT to work on the SpiNNaker Project based in the School of Computer Science. SpiNNaker aims to build and operate a 1-million-core low-powered computer with networking between the cores akin to the connectivity of the brain.
The hardware has now been installed and Andrew is now leading the software development team who are writing software to make it easy to use such a novel parallel system. The software will allow users to describe their task as a graph with the vertices as the processes to be run and the edges the communication between the processes, and then mapping this on to the resources of the machine. This is further abstracted into software that allows users to describe “Spiking Neural Networks” to be run on the machine.
The SpiNNaker Team have recently described the history and development of SpiNNaker in a newly published book SpiNNaker: A Spiking Neural Network Architecture (available for free as a PDF). The book describes the development process from conception; though the designing of the chip and the building of the machine; the previously mentioned software in the chapter “Stacks of Software Stacks”, in which Andrew and Donal were personally involved, along with the others in the SpiNNaker software team; the applications run using the software; deep neural networks; learning neural network models run on SpiNNaker; and finally a look at the ongoing development of the next-generation SpiNNaker-2 chip.
Although the software team hadn’t written a book chapter before, they had written a few papers on the software, so a lot of the work involved combining the content of these papers together in a coherent manner, which made the process much simpler!
Work continues on the SpiNNaker software as part of the Human Brain Project for the next 3 years. Work on the SpiNNaker-2 chip is also underway in partnership with the University of Dresden, who have obtained funding to build the second-generation machine. The SpiNNaker Team are expecting to support this with the continued development of the software stack, adding the adaptations required to allow the software to work on both the current generation and next generation systems.