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 operation is parallel or serial, respectively, these technologies impose different constraints. Analog computation constrains the number of distinct ion-channel populations (subcellular compartments) that can be simulated—unlike digital computation, which simply takes longer to run bigger simulations. Digital communication constrains the rate (bandwidth) at which synapses can be activated—unlike analog communication, which simply sums additional inputs onto the same line. Working within these constraints, Neurogrid achieves its goal of simulating multiple cortical areas in real-time through the following choices.
Each with 256x256 silicon neurons in 11.9x13.9 sq-mm, sixteen Neurocores will be assembled in a tree (insert) to build Neurogrid, connected by 100M spike/sec links [Rodrigo Alvarez 2009].
Neurogrid simulates one million neurons and six billion synapses by using two subcellular compartments (per cell) to minimize ion-channel population count and by using local analog communication to minimize bandwidth, choices motivated by cortical studies. Interactions between projections that terminate in different layers of the cortex 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. And, as expected from cortical areas' map-like organization, input are pooled locally, with axonal ramifications making synapses onto cells within a fixed radius (patchy arbors).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!
Three routing fabrics implement the cortex's horizontal projections, vertical connections, and patchy arbors by relaying spikes to neighboring Neurocores in a binary tree, following a programmable route. An off-chip RAM, at the tree's root, softwires horizontal projections, allowing an axon to target arbitrary locations (10, on average) on one or more Neurocores. An on-chip RAM, in each Neurocore, softwires vertical connections, allowing the axon to send collaterals to corresponding locations (6, on average) on that Neurocore and its daughters—a cortical column. It also specifies the synaptic connection's anatomical location (dendrite or soma) and physiological properties (excitatory or inhibitory). Finally, wires running between neighboring neurons use analog communication to emulate axonal arbors, allowing the axon to invoke postsynaptic potentials (100, on average) within a programmable radius. In this hierarchical fashion, each of Neurogrid's over one million neurons connects to six thousand others.
