In August 2014, IBM researchers led by Dharmendra Modha published in Science a chip called TrueNorth, one of the first large-scale neuromorphic processors - silicon designed to compute the way a brain does. The work came out of DARPA’s SyNAPSE program (Systems of Neuromorphic Adaptive Plastic Scalable Electronics).
TrueNorth packed 5.4 billion transistors into 4,096 cores that together implemented one million programmable spiking neurons and 256 million configurable synapses. Unlike a conventional processor that separates memory from computation and runs to a central clock, TrueNorth interleaved memory and processing in each core and was event-driven: neurons did work only when they spiked, the way biological neurons do. The result was startling efficiency - the chip drew just 70 milliwatts, orders of magnitude less than a CPU or GPU doing comparable neural-network work.
The design deliberately mirrored the brain’s architecture rather than its specific wiring. By co-locating memory and computation and communicating through sparse spikes, it sidestepped the so-called von Neumann bottleneck, the constant shuttling of data between separate memory and processor that limits ordinary computers. Chips could also be tiled together to build larger sheets, like patches of cortex.
TrueNorth did not displace the GPU as the workhorse of mainstream deep learning, which favored dense, clocked arithmetic. But it was a landmark demonstration that brain-inspired, spike-based hardware could be built at scale, and it helped launch the modern neuromorphic-computing effort continued by Intel’s Loihi and Manchester’s SpiNNaker.