Neuromorphic Hardware
Neuromorphic Computing · Frontiers track
8 min read
In this module
Silicon Brains
The previous lesson covered how biological brains compute: event-driven spikes, plastic synapses, massive parallelism, memory and processing welded together. Now for the chips that try to capture this in silicon. Neuromorphic hardware contains some of the wildest chip designs of the past decade — they rethink, from the ground up, how transistors are organised and what they're doing.
IBM TrueNorth (2014)
IBM's TrueNorth was one of the first large-scale neuromorphic chips, and it's still a landmark. Built under DARPA's SyNAPSE programme, TrueNorth proved that brain-inspired computing could actually live in working silicon.
Architecture
TrueNorth contains 4,096 neurosynaptic cores arranged in a 64x64 grid. Each core contains:
- 256 neurons (configurable neuron models)
- 256 x 256 synapses (65,536 synapses per core)
- A local SRAM for storing synaptic weights and neuron states
Across the entire chip:
- 1 million neurons
- 256 million synapses
- 5.4 billion transistors
- 4,096 parallel cores operating asynchronously
TrueNorth Chip (64 x 64 grid of cores):┌────┬────┬────┬─── ─ ─ ─ ───┐│Core│Core│Core│ ... │ Each core:│0,0 │0,1 │0,2 │ │ 256 neurons├────┼────┼────┼─── ─ ─ ─ ───┤ 256×256 synapses│Core│Core│Core│ ... ││1,0 │1,1 │1,2 │ │├────┼────┼────┼─── ─ ─ ─ ───┤│ │ │ │ ││ . │ . │ . │ ... │ 64 × 64 = 4,096 cores total│ . │ . │ . │ │└────┴────┴────┴─── ─ ─ ─ ───┘Key Design Principles
No shared resources: Each core stands alone with its own neurons, synapses, and memory. No shared buses, no caches, no centralised controllers. Contention vanishes; parallelism explodes.
Event-driven operation: Cores talk to each other by sending spike packets. A core receiving no spikes does no computation and burns near-zero dynamic power.
Deterministic operation: TrueNorth is asynchronous internally but uses a tick-based execution model on top. Every tick, all cores process their input spikes, update neuron state, and emit output spikes. The whole thing is reproducible and debuggable.
Ultra-low power: TrueNorth pulls just 65 milliwatts in typical operation — less than a hearing aid battery. That's roughly 10,000x more energy-efficient per synaptic op than a conventional processor running an equivalent neural network simulation.
TrueNorth's miserly power draw comes from three design moves: event-driven compute (idle cores consume next to nothing), co-located memory and compute (synaptic weights live in SRAM inside each core, so no data movement), and 1-bit synaptic weights (drastically cheaper to store and use). You pay for it in precision, but for most pattern-recognition work that's fine.
Limitations
The 1-bit weight cap (on or off) is restrictive — it kneecaps the kinds of networks you can run. There's also no on-chip learning. Weights have to be trained offline on a conventional machine and then burned in. The neuron model is the simple leaky integrate-and-fire variety, which can't reproduce richer biological behaviour.
Intel Loihi and Loihi 2
Intel's Loihi (2017) and Loihi 2 (2021) represent the current state of the art in neuromorphic hardware, addressing many of TrueNorth's limitations.
Loihi Architecture
The original Loihi packs 128 neuromorphic cores, each simulating up to 1,024 neurons — about 130,000 neurons and 130 million synapses per chip. It improves on TrueNorth in a handful of important ways:
Multi-bit synaptic weights: Loihi supports 1- to 9-bit weights, opening up more expressive network models.
On-chip learning: Programmable learning engines implement STDP and other local rules directly on the chip. The network can learn in real time, no external trainer required.
Programmable neuron models: A microcode engine in each core lets you swap in different neuron dynamics — leaky integrate-and-fire, sure, but also fancier models if you need them.
Hierarchical connectivity: A mesh network ties cores together inside a chip, and multiple chips can be stitched into larger systems.
Loihi 2 Improvements
Loihi 2 (built on Intel 4 process technology) doubles down on programmability:
- 128 neuromorphic cores, each supporting up to 8,192 neurons (up to ~1 million per chip)
- Fully programmable neuron models using a custom microcode ISA, enabling researchers to implement virtually any neuron model
- Graded spikes - spikes can carry a multi-bit payload, not just binary fire/no-fire
- Six embedded Lakemont cores for management and interfacing
- Improved learning engines supporting a wider range of plasticity rules
Hala Point: Scaling Up
In 2024, Intel unveiled Hala Point, a neuromorphic research system containing 1,152 Loihi 2 chips. Hala Point has:
- Over 1.15 billion neurons
- 128 billion synapses
- Capabilities comparable to a large cortical area of the brain (in terms of neuron count, though not biological complexity)
Hala Point fits in a six-rack-unit server chassis and is used by researchers to explore large-scale neuromorphic algorithms.
Learning on-chip is a sharp advantage Loihi has over TrueNorth and most other neuromorphic systems. The chip adapts to new data live, with no detour through a separate trainer to reload weights. That unlocks online learning from sensory streams, adaptive control, and continual-learning scenarios where the environment keeps shifting under you.
SpiNNaker (Spiking Neural Network Architecture)
SpiNNaker, out of the University of Manchester, takes a very different swing at neuromorphic computing. Rather than designing custom neuron circuits, it throws millions of small conventional ARM cores at the problem and simulates spiking networks in software.
Architecture
The SpiNNaker 1 machine (completed in 2018) contains:
- 1 million ARM968 processor cores across 57,600 chips
- 18 cores per chip
- Each core simulates approximately 1,000 neurons in real time
- Total: ~1 billion neurons
The chips are connected by a custom packet-switched network designed for spike communication. Each spike is a small packet (40-72 bits) containing the source neuron ID. The network uses multicast routing so that a single spike packet can be delivered to all target neurons efficiently.
Design Philosophy
SpiNNaker's pitch is flexibility over efficiency. General-purpose ARM cores will simulate any neuron model, any connectivity pattern, any learning rule — bounded only by the software and the cycles the ARMs can spare. That makes it perfect for neuroscientists who want to keep tweaking the model.
The price is power. SpiNNaker burns considerably more than dedicated neuromorphic hardware like TrueNorth or Loihi, because general-purpose processors aren't free.
SpiNNaker 2
SpiNNaker 2, currently in development, replaces the ARM968 cores with ARM Cortex-M4F cores (which include floating-point units) and adds hardware accelerators for common neural operations. It also includes dedicated machine learning accelerators for efficient neural computation.
BrainScaleS: Analogue Neuromorphic Computing
The BrainScaleS system, developed at Heidelberg University, takes yet another approach: it implements neurons and synapses using analogue circuits rather than digital ones.
Analogue Neuron Circuits
In BrainScaleS, the membrane potential of a neuron is represented as an actual analogue voltage on a capacitor. Synaptic inputs are implemented as currents that charge or discharge this capacitor. When the voltage crosses a threshold, a digital spike is generated. The neuron dynamics are governed by the physics of the analogue circuits, not by a digital simulation.
This analogue approach has a striking advantage: it operates in continuous time and can run much faster than biological real time. BrainScaleS 2 neurons operate approximately 1,000 times faster than their biological counterparts, meaning a simulation that would take an hour in biological time completes in about 3.6 seconds.
Trade-offs
Analogue circuits are noisier and less precise than digital ones — manufacturing variation and thermal noise see to that. But that imprecision actually mirrors what biological neurons look like, and neural networks are remarkably tolerant of noise. So analogue can end up both faster and more energy-efficient than digital in this regime.
The headache is calibration. Every analogue neuron circuit behaves slightly differently thanks to fab variation, and you need careful per-device calibration to get consistent behaviour across the chip.
Comparing Neuromorphic Architectures
| Feature | TrueNorth | Loihi 2 | SpiNNaker | BrainScaleS 2 |
|---|---|---|---|---|
| Approach | Digital, custom neuron | Digital, programmable | Software on ARM cores | Analogue circuits |
| Neurons per chip | 1M | ~1M | ~18K (per chip) | ~512 |
| On-chip learning | No | Yes | Yes (software) | Yes |
| Weight precision | 1-bit | 1-9 bits | 32-bit (software) | 6-bit analogue |
| Power per chip | 65 mW | ~1 W | ~1 W per chip | ~1 W |
| Speed vs. biology | Real-time | Real-time | Real-time | 1000x faster |
| Primary use | Inference | Research + deployment | Neuroscience research | Neuroscience research |
Emerging Approaches
Several newer neuromorphic designs are worth mentioning:
Akida (BrainChip): A commercial neuromorphic chip targeting edge AI applications. Akida focuses on event-driven processing of data from event cameras (dynamic vision sensors) and is available as a production chip for embedded systems.
DYNAPs (University of Zurich): Mixed analogue-digital neuromorphic processors designed for ultra-low-power always-on sensing applications.
Tianjic (Tsinghua University): A hybrid chip that can run both conventional deep neural networks and spiking neural networks on the same silicon, providing a bridge between the two paradigms.
The Software Challenge
Hardware on its own won't get you there. Programming neuromorphic systems demands a different mental model and an entirely different toolchain:
- Lava (Intel): Python-based framework for Loihi 2. Gives you abstractions for processes, connections, and learning rules.
- Nengo: Neural engineering framework that compiles high-level network descriptions down to various neuromorphic backends.
- PyNN: Common API for neural network simulators; targets SpiNNaker and BrainScaleS.
- Akida SDK: BrainChip's kit for turning trained deep-learning models into spiking networks.
The programming model itself looks alien. You don't write a sequence of instructions. You describe a network of neurons, set their wiring and learning rules, then let spikes flow through. That mental flip trips up plenty of seasoned programmers.
In the next lesson, we will dive deeper into the computational model that runs on this hardware: spiking neural networks and how they process information through time.