Always-on sensing
Keyword spotting, wake-word detection, and vibration monitoring where a device must listen continuously on a battery for months rather than hours.
What is neuromorphic computing?
Neuromorphic computing builds hardware that processes information the way nervous systems do: many simple units exchanging sparse, event-driven spikes, with memory sitting beside the computation instead of behind a separate memory bus. The point is not biological fidelity. It is that activity-proportional computation uses far less energy than clocking dense arithmetic over data that mostly has not changed.
Spikes instead of tensors
A conventional network passes continuous activation values through every layer on every inference. A spiking neural network encodes information in the timing and sparsity of discrete events, so a unit that has nothing to say stays silent and costs nothing. Information lives in when spikes arrive and how often, not in a floating-point number recomputed each cycle.
That changes the economics of inference. On conventional hardware the dominant cost is usually moving data, not arithmetic. Neuromorphic designs attack that directly by co-locating synaptic weights with the circuits that use them and by only doing work when an event occurs.
How it differs from GPU-based deep learning
Conventional accelerators
- Dense matrix arithmetic on a fixed clock
- Activations shuttled between DRAM and compute on every layer
- Energy scales with total work, whether or not the input changed
- Mature toolchains and near-universal framework support
Neuromorphic hardware
- Sparse, event-driven spikes instead of dense tensors
- Synaptic state held beside the compute that uses it
- Energy scales with activity, so quiet inputs cost almost nothing
- Young toolchains; models usually need conversion or retraining
Neither is a general replacement for the other. Neuromorphic hardware wins where inputs are sparse, power is scarce, and latency matters; dense accelerators remain the right tool for training and for batch throughput.
Where it is actually used
Event-camera vision
Event cameras already emit sparse per-pixel changes, which maps naturally onto spiking hardware and avoids reconstructing full frames just to discard them.
Robotic control loops
Tight sensorimotor loops benefit from low, predictable latency and from computation that stays near the sensor rather than crossing a bus.
Edge anomaly detection
Systems that are mostly idle and occasionally interesting suit a substrate whose power draw follows the data rather than the clock.
What is still unsolved
- Training: most accurate spiking models are still converted from conventional networks rather than trained natively, and conversion costs accuracy.
- Benchmarking: energy figures are frequently quoted without a like-for-like baseline, timestep budget, or accuracy point, which makes published numbers hard to compare.
- Toolchains: mapping a model to a specific substrate remains hardware-specific work, so results rarely transfer between platforms.
- Analog variability: device mismatch and drift mean the same network can behave differently across dies and across temperature.
Common questions
- Is neuromorphic computing the same as a spiking neural network?
- No. A spiking neural network is a model; neuromorphic hardware is a substrate designed to run that class of model efficiently. Spiking models can be simulated on a GPU, but the energy advantage largely disappears when you do.
- Can I run an existing model on neuromorphic hardware?
- Usually only after conversion, and conversion trades accuracy for sparsity. Reporting the accuracy point alongside any energy figure is what makes such a result interpretable.
- How should energy results be reported?
- Joules per inference, at a stated accuracy, with the timestep budget and a like-for-like baseline on conventional hardware. Anything less is not comparable across papers.
Run your workload on our testbed
The centre operates an analog neuromorphic testbed with per-core energy measurement and matched baselines on conventional edge hardware. External researchers can propose a workload and run it with the Neural Architecture group.