The Board:
Neurogrid
Inspired by GRAPE-6, a $60K supercomputer that has revolutionized astrophysics, Neurogrid provides an affordable option for brain simulations. It uses analog computation to emulate ion-channel activity and uses digital communication to softwire structured connectivity patterns. Because their operation is parallel or serial, respectively, these technologies impose different constraints. Analog computation constrains the number of distinct ion-channel populations that can be simulated—unlike digital computation, which simply takes longer to run bigger simulations. Digital communication constrains the number of synaptic connections that can be activated per second—unlike analog communication, which simply sums additional inputs onto the same wire. Working within these constraints, Neurogrid achieves its goal of simulating multiple cortical areas in real-time by making the following judicious choices.
Neurogrid will have sixteen Neurocores, each with 256x256 silicon neurons in 11.9x13.9 sq-mm, interconnected by 80M spike/sec links (insert). An off-chip RAM (at the tree's root) and an on-chip RAM (in each Neurocore) softwire horizontal and vertical cortical connections, respectively [Rodrigo Alvarez 2009].
Neurogrid simulates one million neurons by using two subcellular compartments (per neuron), a choice motivated by cortical studies. Nonlinear interactions between projections that terminate in distinct cortical layers have been replicated in a pyramidal-cell model with just two compartments. Furthermore, varying their electrical coupling replicates the firing patterns of various pyramidal-cell types. Using the smallest number of compartments that captures these behaviors lets us minimize the number of distinct ion-channel populations that need to be simulated.Recent breakthroughs in neuromorphic engineering make it possible to combine analog's real-time operation with digital's programmability, reaping the best of both worlds.
Rivaling Blue Gene's performance, Neurogrid will simulate a million neurons in real-time, while consuming a million times less energy, one watt instead of a megawatt!
Neurogrid simulates six billion synaptic connections by using local analog communication, another choice motivated by cortical studies. Cortical axons synapse profusely in a local area, course along for a while, then do it again. Thus, nearby neurons receive inputs from largely the same axons, as expected from the map-like organization of cortical areas. Local wires running between neighboring silicon neurons emulate these patches, invoking postsynaptic potentials within a programmable radius. Using a patch radius of 6 lets us increase the number of synaptic connections a hundredfold—from 600 million to six billion—without increasing digital communication.
