IBM and the US Air Force have built a brain-inspired supercomputing system with the equivalent of 64 million artificial neurons and 16 billion synapses. The system is designed to operate as close to a biological brain as possible.
The system contains 64 of IBM’s TrueNorth chips, fits inside a standard server rack and will be able to scale to half a billion artificial neurons with its current architecture. The TrueNorth chips are different from CPUs in that each core can operate in parallel and without a clock. The system is inherently resilient, If one core stops working the rest can continue operating without interruption.
Just like a biological brain the system is designed for extreme energy efficiency, It requires only 10 watts to power all 64 million neurons and 16 billion synapses. The system isn’t yet as efficient as a biological system but it is getting close, for comparison the human brain contains 100 billion neurons and uses around 20 watts.
The ultra-low power usage will enable vastly more capable mobile artificial intelligence systems for use in self-driving cars, smartphones, and aircraft which is the Air Force’s main interest. IBM has made impressive progress over the last six years with its TrueNorth chip architecture, far outstripping the pace of Moore’s Law.
At a lecture I listened to last year the head of IBM’s Brain-Inspired Computing division, Dharmendra Modha spoke about his ultimate goal of a human brain equivalent in a two liter box. By 2020 Dr. Modha believes his team will have built a 10 billion neuron equivalent system that can fit in a two liter box and require only 20 watts to operate, that would give them a desktop system with roughly 10% the compute power of the human brain that could easily run on a smartphone battery. Dr. Modha also said that in order to achieve this goal by 2020 he only needs 7nm chips, the FABs for which are already under construction and will be producing 7nm chips in 2018, two years ahead of Dr. Modha’s schedule. And as I’ve written about previously IBM has demonstrated that 5nm chips are achievable.
Six years ago IBM’s TrueNorth Neurosynaptic System contained only 256 neurons per system and today it contains 64 million, that’s an annual eight fold increase. If Dr. Modha’s team can maintain this rate of increasing artificial neurons in the system then they are well on their way to building a system with 10 billion artificial neurons by 2020, and a human brain equivalent of 100 billion artificial neurons before 2025.