If you don’t know what the neocortex is, it is the ripply outside of the brain that all mammals have and that is considered the center of intelligence and symbolic processing; you might know it as the cerebrum. Professor Henry Markram at EPFL, according to IBM, has the best understanding of the chemistry of the neocortex, and having spent 10 years building up data on the chemistry of the neocortex, he now wants to simulate a 10,000-neuron column of the neocortex.
To do so on the Lintel cluster that EPFL currently has is not possible, and both IBM Research and EPFL believe that the unique architecture of the Linux-based Blue Gene supercomputer will fit well with the simulation software that EPFL will need to create to simulate this neocortex section. So, to that end, EPFL is acquiring a four-rack Blue Gene system rated at 22.8 teraflops to begin its initial simulations; EPFL has an option to add another two racks if it needs more processing capacity. Each Blue Gene rack lists for about $2.5m, according to Tilak Agerwala, vice president of systems at IBM Research and the person who is steering the Blue Gene project into its commercialized phase.
The ultimate goal of Markham’s research, says Charles Peck, the lead researcher on the IBM side of the joint project with EPFL, is to refine the initial neocortex simulation to the point where it can be accurately described with simpler algorithms than the initial ones IBM and EPFL will create such that the entire human brain can be simulated at an electrochemical level. To do so might take as much as 100 to 1,000 petaflops – that’s 100,000 to 1 million teraflops – of computing power, says Peck, which is well out of reach of any of the technologies we expect to have before 10 to 15 years from now.
At the high end of that range, such a machine by my math would eat up about 165 acres of space using current Blue Gene technology and would consume 4.7 gigawatts of power. Your brain does the actual processing in something the size of a melon, with minimal heat, and can be powered for a few hours on a candy bar. Amazing, isn’t it? Clearly, any full-brain simulations will only be possible if the algorithms can be refined to accurately describe brain chemistry and not take up as much processing power as they initially will.
As part of the deal, IBM’s Zurich lab is getting to use some of the excess capacity in the Blue Gene machine to help it design carbon nanotubes and other potential future semiconductor technologies. Other researchers at EPFL will also use the machine to study protein folding as it relates to mad cow, Creutzfeldt-Jakob, and related brain diseases. The initial Blue Gene machine at EPFL will have about 10 times the peak processing power of its current Lintel cluster.
