Intel announces Loihi, a self-learning neuromorphic chip

On 25 September 2017, Intel Labs announced Loihi, an experimental neuromorphic chip built to run spiking neural networks and to learn on the chip itself. Where IBM’s earlier TrueNorth ran fixed, pre-trained networks, Loihi’s headline feature was that it could adapt while operating, adjusting its synapses from feedback without a separate training phase on another machine.

The chip integrated 128 neuromorphic cores plus a few conventional x86 cores, fabricated on Intel’s 14-nanometer process. It supported up to about 130,000 spiking neurons and 130 million synapses. Crucially, it implemented learning rules based on spike timing - the brain-inspired idea that connections strengthen or weaken depending on the precise timing of the spikes passing through them - so the network could continue learning in place, asynchronously and without a global clock.

Loihi was a research platform, not a product. Intel distributed it to universities and labs through a research community, and used it to explore tasks where event-driven, low-power computation has an edge, such as gesture recognition, smell detection, and certain optimization and search problems. A second-generation Loihi 2 followed, and Intel later assembled many chips into large neuromorphic systems.

The significance, alongside TrueNorth and SpiNNaker, was to keep alive a distinct approach to AI hardware. While the deep-learning boom rode dense GPU arithmetic, neuromorphic chips bet that copying the brain’s sparse, spike-based, memory-near-compute style could one day deliver intelligence at a fraction of the energy.