Neuromorphic Chips

Intel Loihi 2 · IBM NorthPole · BrainChip Akida · SynSense · SpiNNaker2 · Darwin 3

Neuromorphic chips compute only when a neuron spikes, which saves energy on sparse sensor input and little else. The largest system is now Chinese: Zhejiang University's Darwin Monkey reached about 2 billion neurons in August 2025, ahead of Intel's 1.15-billion-neuron Hala Point at Sandia, while BrainChip and SynSense are the only vendors selling chips.

The largest neuromorphic systems by neuron count, log scale: Zhejiang University's Darwin Monkey, about 2 billion neurons on 960 Darwin 3 chips (2025); Intel's Hala Point at Sandia, 1.15 billion neurons on 1,152 Loihi 2 chips (2024); SpiNNaker2 at Sandia, about 175 million neurons (2025); and the human brain at about 86 billion for scale. LARGEST NEUROMORPHIC SYSTEMS, NEURONS (LOG SCALE) 10M 100M 1B 10B 100B Darwin Monkey ZHEJIANG UNIV., 2025 2B 960 DARWIN 3 CHIPS Hala Point INTEL, AT SANDIA, 2024 1.15B 1,152 LOIHI 2 CHIPS SpiNNaker2 at Sandia SPINNCLOUD, 2025 175M ARM CORES ON GF 22FDX Human brain FOR SCALE 86B ABOUT 86 BILLION NEURONS EVEN THE LARGEST SYSTEM HOLDS ABOUT 2% OF A HUMAN BRAIN'S NEURONS.
Largest system
Darwin Monkey, ~2B
Intel Hala Point
1.15B neurons
Commercial vendors
BrainChip, SynSense
IBM NorthPole
22B transistors
Akida power
Under 1 mW
Hub
Chips

How it computes

Spiking Chips Save Energy Only on Sparse Input

Neuromorphic chips model neurons that fire spikes and communicate only when something changes, in place of the clocked fetch-execute loop of a CPU or GPU.

A clocked processor updates every unit on every clock tick whether or not the input changed; a spiking chip does work only when a neuron fires, so a sparse input costs a few events instead of thousands of operations. SAME INPUT, TWO WAYS TO COMPUTE Clocked (GPU, CPU) EVERY UNIT WORKS ON EVERY TICK: 24 OF 24 SLOTS SPEND ENERGY Event-driven (spiking) WORK HAPPENS ONLY WHEN A NEURON SPIKES: 4 EVENTS The saving holds only on sparse input: event cameras, wake words, motor feedback. Dense work such as transformer inference keeps every unit busy; GPUs win there.

The chips run on standard CMOS: Loihi 2 on Intel 4, NorthPole on a 12nm process, Akida on TSMC 28nm and GlobalFoundries 22FDX, SpiNNaker2 on GlobalFoundries 22FDX. Everything neuromorphic is in the circuit design, so supply risk sits in research funding and commercialization, not wafer allocation or packaging.

Platforms

Eight Platforms, Two of Them Commercial

  • Corporate research
  • Commercial or Chinese program
  • Academic
Neuromorphic platforms
PlatformChip and processScaleStatus
Intel Loihi 2Loihi 2 on Intel 41 million neurons a chip; Hala Point system: 1,152 chips, 1.15 billion neuronsResearch only; Hala Point delivered to Sandia National Laboratories in 2024; not sold
IBM NorthPole12nm; 22 billion transistors on an 800mm2 die; 256 cores with all weights in on-chip memoryInference chip, not spikingResearch prototype (Science, October 2023); IBM reports far better energy per frame than GPUs on ResNet-50
BrainChip AkidaAKD1000 on TSMC 28nm; AKD1500 taped out on GlobalFoundries 22FDX (2023); Akida IP for licensingSub-milliwatt event detectionCommercial; listed on the ASX; edge sensing, defense and industrial pilots
SynSenseSpeck (vision), Xylo (audio), Dynap-CNNEvent-camera and audio co-processorsCommercial; headquartered in Zurich with operations in Chengdu; spun out of ETH Zurich's Institute of Neuroinformatics
SpiNNaker2GlobalFoundries 22FDX; 152 Arm Cortex-M4F cores a chipSandia system, about 175 million neurons (2025)Commercialized by SpiNNcloud Systems (Dresden), spun out of TU Dresden
BrainScaleS-265nm mixed-signal CMOS; 512 analog neurons a chipRuns about 1,000 times faster than biological timeAcademic research at Heidelberg University
Darwin 3 (China)Zhejiang University and Zhejiang LabDarwin Monkey: 960 chips, about 2 billion neurons, over 100 billion synapses (August 2025)Research system; the largest neuromorphic computer announced
Tianjic (China)Tsinghua University hybrid ANN/SNN chipSelf-driving bicycle demonstration (Nature, 2019)Research

China

China Holds the Largest System

Zhejiang University and Zhejiang Lab's Darwin Monkey, also called Wukong, wires 960 Darwin 3 chips into about 2 billion neurons and over 100 billion synapses, passing Intel's 1.15-billion-neuron Hala Point when it was announced in August 2025. Tsinghua's Tianjic hybrid chip and SynSense's Chengdu operation give China both research scale and a commercial foothold, and neuromorphic chips sit outside the export controls aimed at AI accelerators because they run on mature or mid-range nodes.

Adoption

Software, Not Silicon, Holds Neuromorphic Back

Transformers, CNNs and diffusion models are trained with backpropagation on continuous values. Spiking networks use discrete events and local learning rules that standard backpropagation does not handle, so existing models must be converted, usually losing accuracy, or rebuilt. Each GPU generation also narrows the energy gap neuromorphic chips set out to close.

IBM NorthPole sidesteps the problem by running ordinary neural network inference with all weights on chip, which makes it the most direct route for neuromorphic-style hardware to reach products.

Where each platform is used, and what stops it
PlatformWhere it is usedBarrier
Intel Loihi 2Optimization, sparse coding and robot control researchNo product; spiking models need specialist programming
IBM NorthPoleEdge inference prototypeNo commercial timeline announced
BrainChip AkidaAlways-on anomaly detection, defense sensors, IoTSmall volumes; model conversion limits; competes with DSPs and NPUs
SynSense SpeckEvent-camera vision for robots and gesture detectionEvent cameras are a niche; the whole sensor stack has to change

Robots

The Strongest Case Is Always-On Robot Sensing

A humanoid carries dozens of sensors that mostly report nothing new. Event cameras paired with SynSense Speck, or Akida running anomaly detection below a milliwatt, could watch them without waking the main compute. The humanoid power budget is on Humanoid Semiconductor Stack; the vehicle and robot demand side is on ElectronsX.

Record

Page Record

Covers
Spiking versus clocked compute, neuromorphic platforms, China's Darwin program, adoption barriers, robot sensing
Sources
Intel Hala Point release (April 2024); IBM NorthPole in Science (October 2023); BrainChip AKD1500 release (2023); Sandia and SpiNNcloud (June 2025); Zhejiang University (August and September 2025); Tianjic in Nature (2019)
Last verified
October 2026