Skip to content
ContentLora

    Tip: press / anywhere to search.

    concept

    Neuromorphic computing

    Also known as brain-inspired computing, spiking neural network hardware, Loihi, Hala Point

    Neuromorphic computing applies principles of neuroscience to chip design, typically using event-driven spiking neurons with memory and compute side by side.[1][2] The largest system, Intel's 1.15-billion-neuron Hala Point, remains a research prototype as of 2026, and a 2025 Nature review set out what is needed to scale the field.[3][4]

    Editor reviewedUpdated Next-gen computing hardwareComputingSemiconductors
    Key facts

    Neuromorphic computing is a field that applies principles of neuroscience to computing systems to mimic the brain’s structure and function.[1] Its main promise is energy efficiency, and its authors present it partly as a response to the fast-rising electricity use of AI.[5]

    How it differs from conventional chips

    In an ordinary computer, the processor and memory are separate, and data must constantly move between them.[6] The brain works differently: neurons store and process information in the same place and fire only when something happens. Neuromorphic chips copy this. Intel’s Hala Point, for example, uses “spiking” neurons that send signals only when needed, with memory and computing built together.[2]

    Most neuromorphic processors implement spiking neural networks (SNNs): asynchronous, event-driven neurons that communicate by spikes, with synaptic state stored next to compute.[2] The 2025 Nature roadmap recommends brain-like hierarchy: dense local connectivity within neurosynaptic cores combined with sparse global connectivity.[7]

    Landmark systems

    Intel Hala Point (2024). Announced April 17, 2024, Hala Point packs 1,152 Loihi 2 processors, made on the Intel 4 process, into a system with 1.15 billion neurons.[8] It draws at most 2,600 watts.[9] Intel says it reaches 15 trillion 8-bit operations per second per watt on conventional deep neural networks.[10] It was initially deployed at Sandia National Laboratories and is a research prototype.[3]

    IBM NorthPole. NorthPole is a brain-inspired inference chip made on a 12nm process with 22 billion transistors, which keeps its memory on-chip rather than relying on external DRAM.[11] In September 2024 IBM reported that NorthPole ran a 3-billion-parameter language model at under one millisecond per token with 72.7 times the energy efficiency of the next-lowest-latency GPU.[12] In a 2U server with 16 NorthPole chips, IBM reported 28,356 tokens per second on that model, with latency 46.9 times lower than the next most energy-efficient GPU.[13] IBM attributes the result to placing memory and processing together on the chip, sidestepping the von Neumann bottleneck; it notes that processor efficiency has tripled every two years while memory bandwidth grows about half as fast.[14]

    The 2025 roadmap

    In January 2025, 23 researchers published a review in Nature setting out how neuromorphic computing could scale up.[4] They argue for exploiting sparsity and hierarchy as the brain does.[7] Sparsity is modelled on brain development, in which many connections form and most are then pruned, which the authors say could make systems far more compact and energy-efficient.[15] They do not expect one winning design; instead they foresee a range of neuromorphic hardware matched to different applications.[16] Two of the authors also secured a $4 million National Science Foundation grant to launch THOR, a network that gives researchers access to open neuromorphic hardware and tools.[17] Their framing ties the field’s relevance to AI’s growing electricity demand, which they describe as projected to double by 2026.[5]

    Open questions

    As of October 2026, the largest neuromorphic system is still a research prototype, not a commercial product.[3] Much of the efficiency evidence comes from the companies building the chips, and comparisons with GPUs depend heavily on the workload chosen.[10][12] Whether spiking hardware can compete on mainstream AI workloads is discussed in the post-Moore hardware debate. Mainstream AI accelerators such as NVIDIA‘s GPUs face the same memory-bandwidth limit that NorthPole’s on-chip memory is designed to avoid.[18][14] The related approach of analog in-memory computing attacks the same data-movement problem with different physics.[6]

    Questions readers ask

    What makes a chip neuromorphic?

    It applies neuroscience principles to mimic the brain's structure and function; systems such as Hala Point use asynchronous, event-based spiking neural networks.[1][2]

    How big is the largest neuromorphic computer?

    Intel's Hala Point has 1.15 billion neurons on 1,152 Loihi 2 processors and draws at most 2,600 watts.[8][9]

    Can I buy a neuromorphic computer?

    Not Hala Point; Intel describes it as a research prototype, initially deployed at Sandia National Laboratories.[3]

    How efficient is IBM's NorthPole?

    IBM reported 72.7 times the energy efficiency of the next-lowest-latency GPU on a 3-billion-parameter language model.[12]

    Sources

    Each numbered claim is a statement we checked against the sources listed with it. Status shows how well established it is.

    1. [1]

      Neuromorphic computing applies principles of neuroscience to computing systems to mimic the brain's structure and function. confirmedas of 2025-01-23

    2. [2]

      Hala Point runs asynchronous, event-based spiking neural networks with memory and computing integrated. confirmedas of 2024-04-17

    3. [3]

      Hala Point was initially deployed at Sandia National Laboratories and is a research prototype rather than a commercial product. confirmedas of 2024-04-17

    4. [4]

      A review by 23 researchers published in Nature on January 22, 2025 set out a roadmap for scaling neuromorphic computing. confirmedas of 2025-01-23

    5. [5]

      The review's authors present neuromorphic computing as a response to AI's electricity consumption, which they say is projected to double by 2026. reportedas of 2025-01-23· interpretation

    6. [6]

      IBM researchers describe the cost of conventional AI inference as the time and energy spent moving model weights between memory and processors. confirmedas of 2025-01-29

    7. [7]

      The review's authors recommend brain-like designs combining dense local connections within cores with sparse global connectivity. confirmedas of 2025-01-23

    8. [8]

      Intel's Hala Point, announced April 17, 2024, contains 1.15 billion neurons on 1,152 Loihi 2 processors made on the Intel 4 process. confirmedas of 2024-04-17

    9. [9]

      Hala Point consumes a maximum of 2,600 watts. confirmedas of 2024-04-17

    10. [10]

      Intel says Hala Point reaches 15 trillion 8-bit operations per second per watt when running conventional deep neural networks. confirmedas of 2024-04-17

    11. [11]

      IBM's NorthPole inference chip is made on a 12nm process with 22 billion transistors in 795 square millimetres and keeps its memory on-chip. confirmedas of 2024-09-26

    12. [12]

      IBM reported in September 2024 that NorthPole ran a 3-billion-parameter language model at under 1 millisecond per token with 72.7 times the energy efficiency of the next-lowest-latency GPU. confirmedas of 2024-09-26

    13. [13]

      IBM reported that a 2U server with 16 NorthPole chips reached 28,356 tokens per second on its 3-billion-parameter test model, with latency 46.9 times lower than the next most energy-efficient GPU. confirmedas of 2024-09-26

    14. [14]

      IBM says NorthPole avoids the von Neumann bottleneck by placing memory and processing together on the chip, noting that processor efficiency has tripled every two years while memory bandwidth grows about half as fast. confirmedas of 2024-09-26

    15. [15]

      The roadmap highlights sparsity as a key feature to emulate, modelled on how the brain forms many connections and then prunes most of them. confirmedas of 2025-01-23

    16. [16]

      The 2025 Nature roadmap's authors do not expect a single neuromorphic design at scale, but a range of hardware suited to different applications. confirmedas of 2025-01-23

    17. [17]

      Two of the roadmap's authors secured a $4 million National Science Foundation grant to launch THOR, a research network giving access to open neuromorphic hardware and tools. confirmedas of 2025-01-23

    18. [18]

      The same analysis argues that memory bandwidth, rather than compute, has become the primary bottleneck for AI workloads, especially when serving models. confirmedas of 2024-03-21

      • AI and Memory Wall · arXiv (published in IEEE Micro) · 2024-03-21 · Abstract (retrieved 2026-10-10)
    Revision history (2)
    1. Page created.
    2. Expanded the roadmap section (sparsity, no single design, THOR network) and NorthPole results (throughput, on-chip memory).

    Created Oct 10, 2026. Last reviewed by an editor on Oct 10, 2026. Next scheduled review: Jan 10, 2027.

    Cite this page

    "Neuromorphic computing." ContentLora, updated Oct 10, 2026. https://contentlora.com/wiki/neuromorphic-computing

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.

    Keep exploring

    Connections
    Analog in-memory co… Shares goal withIntel Research system built by
    ☰ Contents