July 31, 2012 at 6:07 PM ET
A pioneer of early computer tech is now leading an effort to create the best simulation of the brain yet, by connecting a million individual processors together and throwing out a few fundamental rules of computing along the way.
Simulating the brain is an incredibly difficult task; its billions of neurons form trillions of connections and send quadrillions of signals every second. Even the fastest supercomputers on Earth would be unable to keep track of all that. Steve Furber knows this; he was one of the creators of the ARM processor architecture that powers nearly every mobile phone in the world today, and is well aware of the advantages and shortcomings of modern chips.
So, knowing that even something like IBM's Deep Blue can't properly simulate even a part of the brain, Furber decided to take a different tack. The device being built by him and his team is called the Spiking Neural Network Architecture, or SpiNNaker, and once complete it will be the closest analogue to the brain ever created.
When it's finished in 2013, SpiNNaker will have 1,200 circuit boards; each board will have 48 low-power ARM9 processors on it; each processor will have 18 cores; each core will simulate dozens of neurons. Furber estimates that the system should be able to simulate a billion neurons, and in a fashion that is far more brainlike than previous attempts.
Almost every computer is controlled by a central clock that synchronizes all the transistors and gates and every other microscopic feature that makes up a modern CPU. But the brain has no such central clock, and neither does SpiNNaker.
This means that signals will arrive and be sent off without any kind of synchrony, interrupting one another and changing outputs based on millions of tiny random variations. It sounds like chaos, and for things like precise mathematical calculations, it is. But for "fuzzy" calculations like, say, when to let go of a ball you're throwing or what word best ends a sentence, it's just fine — after all, your brain does all those things without calculating to the tenth decimal.
In other words, the SpiNNaker isn't just a simulation of a brain, the way it might work for a plane in a flight simulator. It actually works the way the brain works, or at least much more so than existing computers. Furber believes that this research will not only advance brain science, but may also improve the way computers deal with large problems or thousands of tiny ones. His description of previous and future systems at IEEE Spectrum has much more detail and history.
Testing of the first boards began in May of 2011, and they expect to complete construction near the end of 2013. The SpiNNaker project was made possible by a £5 million (around $8 million) grant from the UK's Engineering and Physical Sciences group, and involves a number of universities and industry partners. You can track its progress at the SpiNNaker homepage.
Devin Coldewey is acontributing writer for NBC News Digital. His personal website iscoldewey.cc.