N2

N2

Brain-inspired. World-ready.

Enquire Read the paper

Individual transistors.

Billions of switches.

AND
AND
AND
AND
AND
AND
AND
AND
AND
AND
AND
AND

Logic gates.

The building blocks.

Neurons 1,024
Synapses 131K
Learning On-chip
Models 5 types

One core. 1,024 spiking neurons.

Five neuron models. Programmable everything.

128 cores. One mesh network.

Every core communicates. Every spike arrives.

N2

Catalyst N2.

131,072 neurons. UK patent filed.

Transistor
Gate
Core
Mesh
Die
5 neuron models
On-chip learning
1000x efficiency
3 dev backends
Hardware validated

1000x more efficient.

N2 only computes when something happens. No event, no energy. While GPUs burn through billions of operations every cycle, N2 sits silent until a spike arrives — then responds in microseconds.

GPU

Always on. Every cycle computes whether it needs to or not.

N2

Event-driven. No spike, no computation, no energy.

Five ways to think.

N2 supports five distinct neuron models, each optimised for different types of computation. Choose the right model for your application, or mix them within a single network.

Standard

Fast, efficient, the default workhorse.

Adaptive

Adjusts its own firing threshold over time.

Complex

Rich spiking patterns. Bursting, chattering.

Graded

Bridges spiking and analog computation.

Resonant

Frequency-selective. Tuned to temporal patterns.

Strengthen Weaken Connections adapt based on timing, activity, and reward

It learns while it runs.

N2 doesn't need to be retrained in the cloud. It learns in real time, on-chip, using the same mechanism your brain uses — strengthening connections that matter, weakening ones that don't.

A fully programmable learning engine lets you define custom rules. The chip adapts to its environment without host intervention.

Python to silicon.

Define networks in Python. Simulate on your laptop. Train on GPU. Deploy to real hardware. The same code runs everywhere — what you test is what you get on silicon.

View SDK on GitHub →

Proven on real hardware.

Not a simulation. Not a model. N2 has been validated on real programmable logic, running real spiking neural networks. Every test passing. Cycle-accurate. Bit-exact.

28/28
Every test. Passing.

Everything. Upgraded.

1 neuron model → 5 neuron models
Fixed learning rules → Fully programmable
Simulation only → Real hardware validation
168 tests → 3,091 tests

How N2 compares.

Catalyst N2 Intel Loihi 2 IBM TrueNorth SpiNNaker
Neuron models 5 3+ 1 Software
On-chip learning Yes Yes — Yes
Programmable logic Yes Yes — —
Real-time learning Yes Yes — Limited
Hardware validated Yes Yes (ASIC) Yes (ASIC) Yes (ASIC)
Open design Yes No No Partial
SDK included Yes Limited — Yes

Technical Specifications

Compute
Cores128
Neurons per core1,024
Total neurons131,072
Synapses per core131,072 (CSR format)
Neuron Models
CUBA LIFCurrent-based leaky integrate-and-fire
Izhikevich4-parameter, rich spiking dynamics
ALIFAdaptive threshold, intrinsic plasticity
Sigma-DeltaRate-coded graded output
Resonate-and-FireOscillatory, frequency-selective
Learning Engine
Plasticity3-factor STDP
Microcode engine16 general-purpose registers
Eligibility traces7-bit, configurable decay
Reward modulationGlobal + per-core reward signals
HomeostasisProportional error, target rate regulation
Weight precision8-bit with stochastic rounding
Communication
RoutingBarrier-synchronized mesh
Graded spikes8-bit payload per spike
Axonal delaysUp to 63 timesteps
Dendrites4-compartment multi-compartment model
SDK
Nameneurocore v3.7.0
LanguagePython
Modules88
Tests3,091
BackendsCPU, GPU, FPGA
Fixed-point modeHardware-accurate quantization
Hardware Validation
Tests passing28/28
PlatformAWS F2 (Xilinx UltraScale+ VU47P)
Clock frequency62.5 MHz
Cores validated16
RTL accuracyCycle-accurate, bit-exact
Intellectual Property
PatentUK patent filed (full architecture)
DesignOpen (SDK on GitHub)

Read the N2 paper

17 pages. Full architecture, SDK, and validation results.

View PDF

Bring N2 to your project.

Licensing, research partnerships, and integration enquiries.

Enquire Read the paper