Neuromorphic Chips: The Next Frontier for AI Performance

Apple’s Tim Cook proclaimed the iPhone X would \\"set the path for technology for the next decade\\" when it debuted. Its integrated neural engine – the first commercial face‑recognition ASIC – foreshadowed the rise of deep neural networks. Today, neuromorphic processors are emerging as the only viable path to sustain AI’s accelerating demands.

Confronted with data‑bandwidth bottlenecks and ever‑increasing computational needs, a recent report by Yole Développement (Lyon, France) argues that sensing and computing must emulate neuro‑biological architectures to move forward.

In an interview with EE Times, Pierre Cambou, Principal Analyst for Imaging at Yole, said neuromorphic sensing and computing could solve most of AI’s current challenges while unlocking new application horizons in the coming decades. \\"Neuromorphic engineering is the next step toward biomimicry and the engine of AI progress\\", he explained.

Why now?

Seventy years after Alan Turing asked \\"Can machines think?\\" and thirty years after Carver Mead introduced neuromorphic engineering, practical brain‑like machines remained elusive. The tide turned with Georgia Tech’s 2006 field‑programmable neural array and MIT’s 2011 chip that emulates adaptive neuronal behavior.

The real breakthrough came when the University of Toronto’s team published \\"ImageNet Classification with Deep Convolutional Neural Networks\\". The AlexNet architecture – an eight‑layer convolutional neural network – enabled the first successful classification of 1.2 million high‑resolution images into 1,000 categories, such as cats and dogs. \\"AlexNet proved that deep learning is more powerful and sparked AI’s meteoric rise\\", Cambou noted.

Most current deep‑learning implementations ride on Moore’s Law, and that has been \\"working fine\\". But as the field grows, demand for high‑performance, low‑power chips will outpace what traditional scaling can deliver. Cambou warned that deep learning \\"will fail\\" if it remains tied to today’s paradigm.

He cites three key obstacles. First, the economics of Moore’s Law: only a handful of players can continue beyond 7 nm, risking a monopoly that stifles innovation. Second, data volumes are outpacing memory technologies, creating a bottleneck. Third, the heat wall: a 7 nm chip delivers roughly one teraflop per watt, yet powering an autonomous vehicle could require around one kilowatt – or one thousand teraflops – a feat current technology cannot meet. Neuromorphic hardware, with its superior energy efficiency, offers a path forward.

Cambou urges a disruptive shift that harnesses emerging memory technologies to boost bandwidth and power efficiency – the essence of neuromorphic computing. \\"AI will keep advancing, and the next leap is neuromorphic\\", he says.

In recent years, multiple initiatives have built silicon neurons that emulate cognition. According to Yole, the neuromorphic market could grow from $69 million in 2024 to $5 billion in 2029 and $21.3 billion in 2034 if technical hurdles are cleared. The ecosystem spans major players such as Samsung, Intel, and SK Hynix, alongside startups like Brainchip, Nepes, Vicarious, and General Vision.

Credit: Yole

Neurons and synapses

Neuromorphic hardware is exiting the lab, driven by collaboration across sensing, computing, and memory sectors. Joint ventures, strategic alliances, and long‑term EU projects like the Human Brain Project are accelerating progress.

Although commercial products are unlikely before 2024, the potential scale is vast. Intel’s 2017 Loihi chip – 130,000 artificial neurons – and its 8‑million‑neuron Pohoiki Beach system illustrate the rapid pace. IBM’s TrueNorth offers 1 million neurons and 256 million synapses, while Brainchip’s Akida delivers 1.2 million neurons and 10 billion synapses.

\"Synapses matter more than neurons\\", Cambou said. \"The next steps involve building applications on the current asynchronous Von Neumann hybrids, followed by RRAM‑based solutions that add billions of synapses\\".

Memory leaders like Micron, Western Digital, and SK Hynix pursue neuromorphic designs, but true breakthroughs may come from focused startups such as Weebit, Robosensing, Knowm, Memry, and Symetrix. Their memristor and phase‑change memory approaches aim to emulate synaptic plasticity, while MRAM is poised to complement neuromorphic stacks.

A neuromorphic sensing ecosystem is also maturing, rooted in Misha Mahowald’s 1991 Silicon Neuron. Today, fewer than ten companies offer event‑based cameras and sensors, including Prophesee, Samsung, Insightness, Inivation, and Celepixel. Frame‑based imagery cannot capture motion as naturally as the human eye; event‑based devices provide real‑time pattern recognition for dynamic scenes.

Cambou noted, \\"Cinema tricks our brains, but computers need the same stream of visual information\\". The same principle applies to auditory, imaging, and behavioral sensors, which are critical for general intelligence.

Yole projects neuromorphic sensing revenues to rise from $43 million in 2024 to $2 billion in 2029 and $4.7 billion in 2034.

Automotive, but not only

While the automotive sector is the most obvious early adopter, neuromorphic sensing and computing will first power industrial and mobile applications – robotics, real‑time perception, and logistics. \\"Deep learning demands massive datasets, but neuromorphic systems learn from a few images or words and capture temporal dynamics\\", Cambou explained.

Within a decade, hybrid in‑memory computing chips could unlock mass‑market autonomous driving, enabling computers to interpret unstructured environments with human‑like acuity.