Computer chips that work like brains

How far are neuromorphic chips from mainstream use?

Neuromorphic chips copy a few features of the brain, such as short electrical spikes and memory placed next to computation. Intel's Hala Point holds 1.15 billion artificial neurons and has been installed at Sandia National Laboratories since early 2024. Energy claims so far come from the developers and host labs, under narrow test conditions. BrainChip began commercial availability and initial production shipments of its Akida AKD1500 processors in June 2026, but no neuromorphic chip is yet in mainstream use.

AI illustration · Illustrative scene, not a photograph of the specific device or experiment.
WHERE WE AREPilots

Close. Early pilots are running in the real world.

Working chips and research systems exist, and initial production shipments of AKD1500 processors in June 2026 went to defence and wearable customers. Energy figures come from developers and host labs under narrow test conditions. On-chip training has been shown in designs and FPGA prototypes, not in measured silicon. No neuromorphic chip is in mainstream consumer or data-centre use, so the stage is pilot rather than scaling.

The tech tree

Everything recorded on Vrulum that this future needs, oldest foundations at the top and the goal at the bottom. Lines show what each step requires.

What it takes: Computer chips that work like brains Unlocked Being researched Locked
  1. 1990Neuromorphic computing chipsResearch and prototypes Being researched.
  1. 2023Neuromorphic prototype chip shown for image classificationDone Unlocked.
  2. 2024Largest neuromorphic system installed for researchDone Unlocked.
  3. 2024Processor design for training spiking networks on the chipIn progress Being researched.
  4. 2025Scientific computing method mapped onto a spiking processorDone Unlocked.
  5. 2025SpiNNaker2 neuromorphic system deployed at a US government labDone Unlocked.
  6. 2026Commercial neuromorphic processors shippingIn progress Being researched.
  7. Date not recordedMeasured energy on fabricated training siliconNot started Locked.
  8. Date not recordedIndependent benchmarks and mature software for production useNot started Locked.
  1. GoalComputer chips that work like brainsGoalRequires: Neuromorphic computing chips, Neuromorphic prototype chip shown for image classification, Largest neuromorphic system installed for research, Scientific computing method mapped onto a spiking processor, SpiNNaker2 neuromorphic system deployed at a US government lab, Processor design for training spiking networks on the chip, Commercial neuromorphic processors shipping, Measured energy on fabricated training silicon, Independent benchmarks and mature software for production use. Locked.

The road so far

4 of 8 steps done, 2 in progress.

  1. Done · Oct 2023Neuromorphic prototype chip shown for image classification

    IBM Research describes NorthPole as a 12 nm chip with 22 billion transistors. On ResNet-50 image classification, it reports 25 times the energy efficiency (per joule) of common 12 nm GPUs and 14 nm CPUs, and lower latency. The article gives no absolute latency or per-watt figures. [7]

  2. Done · Feb 2024Largest neuromorphic system installed for research

    Intel's Hala Point packages 1,152 Loihi 2 processors, with up to 1.15 billion neurons and 128 billion synapses, and draws a maximum of 2,600 watts. Sandia's news article reports that it arrived in early 2024, and its photo caption gives the delivery date as 1 February, without a year. [1][2]

  3. Done · Jan 2025Scientific computing method mapped onto a spiking processor

    A preprint by Theilman and Aimone maps the finite element method onto spiking neurons running on Intel's Loihi 2. Its abstract reports accuracy and scaling comparable to conventional methods for the Poisson equation, and gives no energy or speed figures. The journal version was not checked. [3]

  4. Done · Aug 2025SpiNNaker2 neuromorphic system deployed at a US government lab

    Sandia announced a SpiNNaker2 system, named Braunfels, to explore how neuromorphic computing could affect nuclear deterrence missions. Sandia states that the chips' accelerators run AI and machine learning algorithms more efficiently than GPUs. This is Sandia's own claim, and the page gives no measured figures. [4]

  5. In progress · Nov 2024Processor design for training spiking networks on the chip

    A preprint describes a multi-core design, with separate feedforward, backpropagation and weight-gradient engines in each core, to train spiking networks directly. It reports 1.05 TFLOPS/W at FP16 on a 28 nm process and a 55% to 85% reduction in DRAM access against an NVIDIA A100 during spiking network training. The abstract does not say whether the efficiency figure was measured on silicon or estimated. FPGA demonstrations ran deep spiking network training on 20 cores and federated learning with 5 workers. The journal version was not checked. [5]

  6. In progress · 30 Jun 2026Commercial neuromorphic processors shipping

    BrainChip announced commercial availability and initial production shipments of its Akida AKD1500 processors on 30 June 2026. The chip is made by GlobalFoundries on a 22 nm FD-SOI process. Production quantities are going to defence and wearable customers while the chip completes industrial and military-grade qualification. The release gives no unit counts, order values or revenue. [6]

  7. Not startedMeasured energy on fabricated training silicon

    No fabricated chip with measured power for on-device training appears in the sources reviewed. The abstract reports no measured power, area or cost for a fabricated chip. [5]

  8. Not startedIndependent benchmarks and mature software for production use

    Efficiency figures so far come from the developers or the host labs. The sources reviewed contain no independent comparison of neuromorphic and GPU systems on the same production workload, and no widely used software toolchain for neuromorphic hardware. [1][2]

Why it matters

If the energy savings hold up on real workloads, spiking chips could run some sensing and AI tasks on small batteries, close to the sensor, instead of in data centres. The field still lacks independent, whole-system measurements, so the size of that benefit is not yet known.

What stands in the way

  • Software for the largest systems is still being built. Sandia says it has been developing algorithms to use Hala Point efficiently and applications that can use the full system, because a system of this scale has not existed before.
  • Spiking networks are hard to train on this hardware. The multi-core preprint states that existing neuromorphic architectures cannot directly train spiking neural networks with backpropagation.
  • Headline efficiency figures depend on test conditions. Intel's 15 TOPS/W figure for Hala Point uses 8-bit precision, random-noise input, a network pruned to 10:1 sparsity and neuron models with 10 percent activation, as set out in Intel's test footnote.
  • On-chip training has no measured silicon result in the sources reviewed. The 1.05 TFLOPS/W figure's measurement basis is not stated in its abstract.
  • New chips must pass industrial and military qualification before volume use, and BrainChip has not disclosed unit volumes or revenue for the AKD1500.

Signals to watch

  • Measured power and efficiency from fabricated silicon for on-chip training designs.
  • Independent benchmarks comparing neuromorphic and GPU systems on the same workload, with energy measured for the whole system.
  • BrainChip reporting AKD1500 unit volumes, repeat orders and named products that have passed qualification.
  • Published results from Hala Point or SpiNNaker2 systems on problems that conventional computers were already solving, with energy reported.
  • Toolchains that compile conventional neural networks onto several neuromorphic platforms without hand rewriting.

Sources

  1. company announcement
    Intel Builds World's Largest Neuromorphic System to Enable More Sustainable AI
    Intel Corporation, 17 Apr 2024.
  2. news
    1.15 billion artificial neurons arrive at Sandia
    Neal Singer. Sandia National Laboratories (Sandia Lab News), 18 Apr 2024.
  3. preprint
    Solving Sparse Finite Element Problems on Neuromorphic Hardware
    Bradley H. Theilman, James B. Aimone. arXiv, 17 Jan 2025.
  4. news
    Sandia Deploys SpiNNaker2 Neuromorphic System
    Sandia National Laboratories, 19 Aug 2025.
  5. preprint
    A High Energy-Efficiency Multi-core Neuromorphic Architecture for Deep SNN Training
    Mingjing Li, Huihui Zhou, Xiaofeng Xu et al.. arXiv, 26 Nov 2024.
  6. company announcement
    BrainChip Announces Commercial Availability and Production Shipments of AKD1500 Neuromorphic Processors
    BrainChip, 30 Jun 2026.
  7. company announcement
    A new chip architecture points to faster, more energy-efficient AI
    Mike Murphy. IBM Research, 19 Oct 2023.

Last reviewed 21 Oct 2026.