SpiNNaker is a biologically inspired, massively parallel computing system optimized for modeling and simulating large-scale real-time networks. In this work, we use the SpiNN-5 (SpiNNaker 103) board 1 2, which consists of 48 SpiNNaker chips. Each chip contains 18 ARM cores with a 32 kB ITCM (instruction tightly coupled memory) and a 64 kB DTCM (data tightly coupled memory) per core. Moreover, a 128 MB SDRAM is shared between the 18 cores. To imitate the high connectivity of the brain, the cores are interconnected by an asynchronous Network-on-Chip (NoC) through a multicast packet-routing mechanism. In addition, SpiNN-5 uses three Xilinx Spartan-6 FPGAs for high-speed serial links.
A 100 MB Ethernet controller handles the connection between the SpiNNaker board and the computer. We use it to load data to the SpiNNaker memory to perform a real-time simulation. Furthermore, the sPyNNaker 3 is a software package used to define models in PyNN script 4 and translates models into a suitable form for SpiNNaker.
Usefull Scripts
To facilitate the use of the board, a couple of scripts are available for use:
- Weights extractor
- Activity visualizer
- CSNN-to-Spinn5 weight transfer
- Spinn-5 classification accuracy
In addition, example of use are alos avialable here:
Official Documentation
Software for SpiNNaker:
https://spinnakermanchester.github.io/
PyNN: A Python package for simulator-independent specification of neuronal network models:
https://neuralensemble.org/PyNN/