Slashdot Mirror
SpiNNaker Powers Up World's Largest Supercomputer That Emulates a Human Brain
The world's largest neuromorphic supercomputer, the Spiking Neural Network Architecture (SpiNNaker), was just switched on for the first time yesterday, boasting one million processor cores and the ability to perform 200 trillion actions per second. HotHardware reports: SpiNNaker has been twenty years and nearly $19.5 million in the making. The project was originally supported by the Engineering and Physical Sciences Research Council (EPSRC), but has been most recently funded by the European Human Brain Project. The supercomputer was designed and built by the University of Manchester's School of Computer Science. Construction began in 2006 and the supercomputer was finally turned on yesterday.
SpiNNaker is not the first supercomputer to incorporate one million processor cores, but it is still incredibly unique since it is designed to mimic the human brain. Most computers send information from one point to another through a standard network. SpiNNaker sends small bits of information to thousands of points, similar to how the neurons pass chemicals and electrical signals through the brain. SpiNNaker uses electronic circuits to imitate neurons. SpiNNaker has so far been used to mimic the processing of more isolated brain networks like the cortex. It has also been used to control SpOmnibot, a robot that processes visual information and navigates towards its targets.
SpiNNaker is not the first supercomputer to incorporate one million processor cores, but it is still incredibly unique since it is designed to mimic the human brain. Most computers send information from one point to another through a standard network. SpiNNaker sends small bits of information to thousands of points, similar to how the neurons pass chemicals and electrical signals through the brain. SpiNNaker uses electronic circuits to imitate neurons. SpiNNaker has so far been used to mimic the processing of more isolated brain networks like the cortex. It has also been used to control SpOmnibot, a robot that processes visual information and navigates towards its targets.
Are you assuming one neuron per core? You can have a neural network with many, many units running on a single CPU core. In that case the number of units in in the hundreds to thousands, typically, which means that neuron-to-neuron communication is relatively easy to handle as you can simply use shared memory. The trick with SpiNNaker and similar efforts is being able to marshall communication with more diverse connections and communication, and that gets complex when communications are not within the locality (physically speaking) of a neuron's computational location. An analogy (and one that Wolfram would approve of!) is to look at cellular automata (I did a fair bit of work on these in the past) where again it is possible to relatively easily do synchronous updates for CA that are locally connected, but much more complex for those that include a neighbourhood with less local connections for the update rule (cf. activation function). Sometimes using asynchronous updates can be useful, as long as they are reasonably timely, and you accept some jitter, but if the activation function is appropriate, it is possible for it to cope with the noise that asynchronicity injects. In my work that is where neural networks came in, as hand designing update rules becomes impossible, and you end up needing a neural network to learn the update complex function based on desired state transitions. But then you still need really good organisation of the distributed computing element to make if efficient in terms of computation/time and even more skill to create something that is efficient in terms of computation/energy. If you look at Steve Furber's credentials with ARM, you can see why he's a good lead for the latter, and he has assembled an excellent team working on all aspects of the problem.