Menu

Neuromorphic computing in 2026

Neuromorphic computing in 2026: what is real and what is still a promise

ReviewNovember 18, 2026Senova Systems Research

A field guide to brain-inspired computing: the chips that exist, how spiking networks are trained, where the measured gains are, and the energy accounting trap that makes many claims look better than they are.

What the word means

Neuromorphic computing borrows three ideas from the brain. Compute happens only when something changes (events, or spikes), not on every tick of a clock. Memory sits next to the arithmetic instead of across a bus. And most of the network is quiet most of the time: sparsity is the point, not a side effect.

A January 2025 review in Nature by 23 researchers, led by Dhireesha Kudithipudi, frames the field's next step as scale: systems that work at the size, speed and energy real tasks need. It names sparsity as the brain's defining feature to copy, and it does not expect one winning chip. Different applications will want different hardware.

The hardware that exists today

Intel's Loihi 2 (announced September 2021, made on a pre-production Intel 4 process) has 128 neuromorphic cores and about one million neurons per chip. Its spikes can carry a value (graded spikes) instead of only a 1 or a 0.

Intel's Hala Point (April 2024, at Sandia National Laboratories) puts 1,152 Loihi 2 chips into a six-rack-unit box: 1.15 billion neurons, 128 billion synapses and 140,544 cores, drawing at most 2,600 W. Intel reports up to 20 petaops and, from early results, as much as 15 TOPS/W on deep networks. It is a research prototype, not a product.

SpiNNaker2, from TU Dresden and the University of Manchester, takes a different route: 152 Arm Cortex-M4F cores per chip in 22 nm FDSOI, with a packet router between chips and voltage and frequency that follow how busy the spiking is.

IBM's NorthPole (Science, October 2023) is a digital inference chip that keeps memory and compute together on a 12 nm die. In 2024 IBM reported 72.7 times the energy efficiency of the next-lowest-latency GPU on a 3-billion-parameter language model. The GPUs it was compared with were on 4 to 7 nm processes, so the comparison is not like for like on silicon.

Analog in-memory computing does the multiply where the weights are stored. IBM's 64-core phase-change-memory chip (Nature Electronics, 2023, 14 nm) reached 92.81 % on CIFAR-10 with an effective matrix precision of 3 to 4 bits. Low precision is the price of computing in analog.

At the edge, BrainChip's Akida (2021) was among the first neuromorphic processors sold commercially. Innatera's Pulsar (May 2025) is a microcontroller that pairs a spiking-network engine with a RISC-V core for battery-powered sensors. The energy and latency gains quoted for both are vendor figures we could not check independently.

How spiking networks are trained

Spikes are all-or-nothing, so the usual gradient-based training does not apply directly. Two approaches dominate.

Conversion: train an ordinary network with ReLU activations, then swap each neuron for an integrate-and-fire neuron whose firing rate stands in for the activation. Rueckauer and colleagues (Frontiers in Neuroscience, 2017) converted networks as large as VGG-16 and Inception-v3. On small networks they reported more than halving the operation count for an error rise of a few percentage points. The catch is latency: a firing rate needs time to settle.

Surrogate gradients: train the spiking network directly by replacing the spike's missing derivative with a smooth stand-in during backpropagation. The tutorial by Neftci, Mostafa and Zenke (IEEE Signal Processing Magazine, 2019) is widely credited with reviving the field.

Where the gains are measured

The most useful finding in the field is a negative one. In their 2021 Proceedings of the IEEE survey of Loihi results, Mike Davies and colleagues report that ordinary feedforward deep networks show little benefit on the chip. The large gains in energy and latency come from networks that use what the hardware is good at: recurrence, precise spike timing, plasticity, randomness and sparsity.

Sensors that are events already pair naturally with this. Event cameras report per-pixel brightness changes with microsecond latency and about 140 dB of dynamic range, against about 60 dB for a frame camera (Gallego et al., IEEE TPAMI, 2022).

The clearest robotics result is a drone. In Science Robotics (May 2024), Paredes-Vallés and colleagues at TU Delft flew a drone whose vision and control ran as one spiking network on a Loihi chip at 200 Hz. The network added 7 to 12 mW on top of the chip's 0.94 W idle draw.

You do not need spikes to get the event-driven benefit. Sigma-delta networks (O'Connor and Welling, ICLR 2017) and delta networks (Neil et al., ICML 2017) pass on only changes in activations, so cost follows how much the input changes. Neil and colleagues reported 5.7 times fewer operations on a speech task with negligible accuracy loss, and a 100-fold cost reduction for a driving network's recurrent layers.

The energy accounting trap

Many neuromorphic claims count arithmetic: a spike triggers additions, and an addition costs less than a multiply-accumulate. Mark Horowitz's ISSCC 2014 keynote supplies the usual 45 nm figures: 3.7 pJ for a 32-bit floating-point multiply and 0.9 pJ for an add. The same keynote shows why counting arithmetic misleads. Reading DRAM costs 1 to 2 nJ, a couple of orders of magnitude more than the arithmetic or a cache access (around 10 pJ). Where the weights live, and how often they move, usually decides the energy.

So in our own studies we report operation counts and an arithmetic-only energy estimate, and we say plainly that memory is left out. A real chip measurement is the only way to settle it.

Our view

Neuromorphic hardware is real, but it is still mostly in research labs. The results that hold up share a shape: a sensor that produces events, a task that is quiet most of the time, and a power budget too tight for a GPU. That describes much of the motion and sensing hardware we design: a levitating stage that mostly hovers, a positioner that mostly holds still, a gripper that mostly waits. That is why our studies start there.

Sources

  1. Kudithipudi et al., Neuromorphic computing at scale, Nature 637 (2025)
  2. DatacenterDynamics, Intel unveils Hala Point (April 2024)
  3. The Register, Intel Hala Point (April 2024)
  4. Loihi 2 overview, WikiChip Fuse (2021)
  5. Höppner et al., SpiNNaker2 processing element architecture (arXiv 2103.08392)
  6. SpiNNaker2: a large-scale neuromorphic system (arXiv 2401.04491)
  7. IBM Research, NorthPole LLM inference results (2024)
  8. Le Gallo et al., 64-core PCM in-memory compute chip, Nature Electronics (2023)
  9. Innatera, Pulsar microcontroller (2025)
  10. Rueckauer et al., Conversion of continuous-valued deep networks to efficient event-driven networks, Frontiers in Neuroscience (2017)
  11. Neftci, Mostafa and Zenke, Surrogate gradient learning in spiking neural networks (arXiv 1901.09948)
  12. Davies et al., Advancing neuromorphic computing with Loihi, Proceedings of the IEEE (2021)
  13. Gallego et al., Event-based vision: a survey, IEEE TPAMI (2022)
  14. Paredes-Vallés et al., Fully neuromorphic vision and control for autonomous drone flight, Science Robotics (2024)
  15. O'Connor and Welling, Sigma-delta quantized networks (ICLR 2017)
  16. Neil et al., Delta networks for optimized recurrent network computation (ICML 2017)
  17. Horowitz, Computing's energy problem (and what we can do about it), ISSCC 2014