Neuromorphic computing is the design of computer chips and systems that imitate the structure of biological brains: many simple “neurons” that communicate through brief electrical spikes, with memory and processing located together. Its main promise is running AI with far less energy than conventional processors.
Origins
The engineer Carver Mead at Caltech coined the term in the late 1980s for analog circuits that mimic neural systems. Large digital neuromorphic chips followed: IBM’s TrueNorth (2014) contained one million artificial neurons and consumed about 70 milliwatts, and Intel introduced its Loihi research chips in 2017 and Loihi 2 in 2021. In April 2024 Intel announced Hala Point, a system of 1,152 Loihi 2 chips with 1.15 billion neurons, which it described as the largest of its kind. In Europe, the SpiNNaker (Manchester) and BrainScaleS (Heidelberg) systems were developed with UK and EU funding and later supported by the Human Brain Project.
Why it differs
Conventional computers shuttle data between separate memory and processors, which costs much of their energy (see Von Neumann Architecture). Brains run on about 20 watts by computing where information is stored and by being active only when spikes occur. Neuromorphic chips copy these features, which suits sensing, robotics and always-on devices (see On-Device AI).
Limits
Spiking networks are harder to train than standard neural networks, software tools are immature, and neuromorphic chips have not matched GPUs on the large models that dominate AI. Many researchers therefore see them as complementary rather than as a replacement, while related “in-memory computing” approaches aim to bring some of the same efficiency to mainstream chips (see AI Accelerator).
Why it matters
As AI’s electricity use grows (see AI Energy Use), the brain remains the proof that intelligence can be far more energy-efficient – and neuromorphic computing is one of the most direct attempts to learn from it.
Written with Claude, made by Anthropic, whose models run on GPUs and other accelerators from investors and partners such as Nvidia, Google and Amazon rather than on neuromorphic hardware; IBM, whose TrueNorth chip is described here, is an Anthropic business partner.
Related
Von Neumann Architecture · AI Accelerator · AI Energy Use · On-Device AI · GPU · Neural Network
Sources: Carver Mead: Neuromorphic Electronic Systems (Proceedings of the IEEE) (1990) · Paul A. Merolla et al.: A million spiking-neuron integrated circuit with a scalable communication network and interface (Science) (2014) · Intel: Intel Builds World’s Largest Neuromorphic System to Enable More Sustainable AI (April 17, 2024)
Leave a Reply