SpiNNaker arrives in the Bayes!

Author: Guest blogger
Posted: 6 Sep 2018 | 11:28

The SpiNNaker neuromorphic high-performance computing platform, which aims to run 1% of the human brain in real time, will arrive at EPCC in the Bayes building this year. SpiNNaker is a novel hardware platform due to its massive parallelism, multi-cast communication fabric and low power design.

The SpiNNaker communication fabric takes a very different approach to normal HPC systems. It is designed to be efficient in sending small data packets very quickly to many places, instead of the standard approach of large blocks of data to be sent to a few places periodically. With HPC systems becoming more multi-core systems, a platform such as SpiNNaker is in an interesting position to explore massive parallel designs.

As each SpiNNaker chip is benched at 1W, the entire one-million core machine utilises approx. 140 KW of power, when taking into account the cooling and the FPGAs that each board uses to create the torus-shaped communication fabric. This is a magnitude lower power than HPC systems for the same processor capability, and is a step towards future HPC systems, which need to reduce the power used to operate them in today's culture. Applications that work efficiently on the SpiNNaker machine are likely to also work well when ported back to traditional HPC systems, as SpiNNaker forces the user to think about their problem in a distributed fashion from the get go. This all makes SpiNNaker a very interesting environment to work with.

SpiNNaker at the Bayes

The SpiNNaker machine directly available to users in the Bayes is a three-board toroid, which provides the user with up to 2304 ARM cores to work with as shown in the image here (right). 

Any user who requires more resources can get access to the half million cores currently available from Manchester University, as shown below. This machine will have grown to one million cores by the end of the year.

Current plans for use of the SpiNNaker platform at the Bayes include supporting research into executing MicroPython in a distributed fashion, as well as supporting the standard users in computational neuroscience. Please keep an eye out for Masters projects which will use the SpiNNaker platform for a multitude of uses. 

Previous works within the computational neuroscience community with the SpiNNaker platform includes:

• Simulating a 1mm by 1mm cortical microcircuit [1] over six SpiNNaker boards. This showed that the new SpiNNaker software is currently at par with the NEST supercomputer, which means the SPiNNaker machine can only get better

• Tracking a ball with the ICUB robot [2]. This showed that by exploiting the weak scaling effect that the SpiNNaker machine easily provides, it can outperform a standard CPU with just one SpiNNaker board.

Find out more

If you are interested in learning about SpiNNaker or how to utilise the platform, please contact the SpiNNaker team through the SpiNNaker user mailing list [3] or contact me directly. I will be available at the Bayes for three weeks each month. 

 [1] Full-scale simulation of a cortical microcircuit on SpiNNaker.
[2] ATIS + SpiNNaker: a Fully Event-based Visual Tracking Demonstration.
[3] https://groups.google.com/forum/#!forum/spinnakerusers
[4] http://www.cs.man.ac.uk/~stokesa6/

Further information

http://apt.cs.manchester.ac.uk/projects/SpiNNaker/
http://apt.cs.manchester.ac.uk/ftp/pub/apt/papers/LAP_IEEEDandT_07.pdf

Author

Alan Stokes, School of Computing Science, University of Manchester

 

Blog Archive