The enabler of neuromorphic edge AI

Always-on AI for the physical world, at microwatt budget.

Spiketronix designs ultra-low-power neuromorphic AI chips, and the software to run them, built on non-volatile compute-in-memory.

Explore the technology Contact us
Spiketronix neuromorphic edge AI chip on a circuit board
Why edge intelligence must change

Too far for the cloud. Too costly to idle.

Power

Years on a coin cell, or self-powered.

Rare events should not require a processor and volatile memory to remain active around the clock.

Speed

Decide in microseconds, on the device.

Local inference removes network delay and reacts while the signal still matters.

Safety

Catch the fault before the machine fails.

Continuous local sensing keeps critical data private and surfaces early warning patterns.

Three technologies, one solution

Built around the sparse, temporal nature of the physical world.

Each pillar removes a different source of wasted energy. Together, they make always-on intelligence practical.

Neuromorphic chip: sparse input spikes converge on a processing core

Neuromorphic

Sleeps until something changes.

Spiking neurons stay silent until their input changes, then emit a brief spike — no event, no computation, no energy. Time and sequence are built in, so vibration, sound and biosignals are computed the way the brain does.

Event-driven Temporal Sparse
FeRAM device with a ferroelectric polarization layer holding its state

Non-volatile FeRAM

Remembers the model with power removed.

Once compute is sparse, the real cost is remembering the model through the long gaps between events. FeRAM holds every weight with the power fully off — so the gaps cost nothing, and the chip is instant-on.

GlobalFoundries 22nm FD-SOI Zero standby
Compute-in-memory crossbar array with an active row and column outputs

In-memory compute

Computes where the weights are stored.

A conventional chip shuttles weights to a separate compute unit every operation — and moving data wastes time and energy. Compute-in-memory does the maths where the weights already live, so nothing moves.

Analog in-memory MAC No weight movement Massively parallel
A complete platform

From silicon architecture to trained model and deployment.

The platform isn't a bare accelerator block you're left to wire up. Silicon, model and toolchain are co-designed as one system, so it ships as a finished part you can build a product around.

Plug-and-play hardware

Mixed-signal accelerator SoC

The neuromorphic compute engine and memory system integrated as one chip platform.

Standard digital interfaces

Designed to connect with conventional embedded systems and host controllers.

On-chip calibration and control

Device variation is handled inside the platform rather than pushed onto the customer.

Integration-ready platform

A deployable architecture rather than a custom research demonstrator.

Spiketronix neuromorphic edge AI chip, top view Neuromorphic SoC Ultra-low-power edge AI
Full software stack

Hardware-aware AI training

Models are trained around the actual memory, quantization and compute behaviour.

Configuration SDK and drivers

Standardized configuration and control instead of custom laboratory scripts.

Model deployment toolchain

Moves a trained network into a configured, deployable hardware implementation.

Adaptive behaviour

The deployed model can be calibrated to the device and its operating environment.

Exploded view of the chip stack, from software down to the ferroelectric core
  1. 01
    Hardware-aware SDKTraining and deployment toolchain.
  2. 02
    On-chip controlCalibration and host interface.
  3. 03
    Spiking neuron fabricEvent-driven digital neurons.
  4. 04
    Mixed-signal front-endSensor input, ADCs and sense amps.
  5. 05
    Compute-in-memory crossbarMultiply-accumulate inside the array.
  6. 06
    FeRAM technologyNon-volatile ferroelectric devices, the core.
Inside the chip A hardware-aware software stack drives the spiking neuron fabric and mixed-signal front-end, all sitting on FeRAM crossbars that compute in memory. The full architecture, realised in one chip.
Always-on machine health monitoring

Spiketronix intelligence at the point of sensing.

Spiketronix module monitoring an industrial motor

Industrial equipment often shows subtle changes long before a visible fault or breakdown occurs.

Spiketronix learns the normal behaviour of a machine directly on the device, detects deviations early and triggers an alert — without continuously sending raw sensor data to the cloud.

  • Earlier warningIdentify developing faults before they lead to unplanned downtime, costly maintenance or equipment damage.
  • Years of autonomous operationUltra-low-power processing enables continuous monitoring where cabling, frequent battery replacement or permanent connectivity are impractical.
  • Decisions, not raw dataOnly relevant events and results leave the device, reducing bandwidth, keeping sensitive operational data local and simplifying integration.

Built for always-on sensing.

The same ultra-low-power architecture can be applied wherever continuous local sensing and immediate decisions are required.

  • Keyword spotting
    Keyword spotting

    Recognise spoken commands and wake-words on the device, without streaming audio to the cloud.

  • Structural monitoring
    Structural monitoring

    Identify changes in buildings, infrastructure and mechanical systems.

  • Health and wearables
    Health and wearables

    Recognize relevant physiological events without continuous data transmission.

  • Smart environments
    Smart environments

    Detect changes in occupancy, activity and environmental conditions.

How deployment works

From raw sensor data to a chip that keeps improving.

  1. 01

    Sensor data collection

    Capture vibration, sound or biosignals from the equipment in real conditions.

  2. 02

    Model training

    Train a hardware-aware model matched to the chip's memory and compute.

  3. 03

    Chip programming

    Write the trained weights into on-chip FeRAM — they hold with the power off.

  4. 04

    Sensor integration

    Mount the programmed module onto the equipment; it runs standalone.

  5. 05

    Local anomaly detection

    The chip watches continuously and flags the drift toward failure on-device.

  6. 06

    Model updates

    Field data refines the model and reprograms the chip over its life.

Step 06 loops back to Step 01
The team behind Spiketronix

Deep expertise across device, system and commercialization.

Thomas Kämpfe

Thomas Kämpfe

Ferroelectric devices and neuromorphics; professor at TU Braunschweig and team lead at Fraunhofer IPMS.

Scientific Advisor

thomas.kaempfe@spiketronix.ai
Contact

Get in touch.

We are happy to answer questions and provide more details about Spiketronix. Reach out for collaborations, partnerships or additional information.