Neuromorphic Computing

1. Why We Need Different Hardware

The success of modern artificial intelligence is built almost entirely on GPUs. However, the computational paradigm of GPUs differs fundamentally from that of the brain:

Dimension	GPU	Human Brain
Power consumption	Hundreds to thousands of watts	~20 watts
Computation mode	Synchronous, clock-driven	Asynchronous, event-driven
Memory and compute	Separated (von Neumann architecture)	Unified
Data transfer	Massive data movement, high energy cost	Local computation, minimal data movement
Activation mode	Continuous values (floating-point)	Discrete spikes

The brain achieves perception, motor control, and reasoning capabilities — feats that current AI systems struggle to match even at thousands of watts — using only 20 watts. This order-of-magnitude gap suggests:

The problem may not lie solely in algorithms; the hardware architecture itself is a bottleneck.

Neuromorphic computing was born precisely to bridge this gap — it attempts to mimic the brain's structure and operating principles at the hardware level.

2. Core Design Principles

Neuromorphic chips differ from conventional processors in several fundamental ways:

Event-Driven Computing

Traditional processors operate on a fixed clock cycle, regardless of whether any meaningful computation needs to be performed. Neuromorphic chips activate computation only upon receiving spike signals — no input means no power consumption.

This mirrors neurons in the brain: they remain silent most of the time, firing spikes only when needed.

In-Memory Computing

In the von Neumann architecture, data must be shuttled back and forth between memory and the processor. The energy consumed by this process far exceeds that of the computation itself — this is the von Neumann bottleneck.

Neuromorphic chips co-locate storage (synaptic weights) and computation (neurons), eliminating the overhead of data movement. In the brain, each synapse both stores information (connection strength) and participates in computation (signal transmission) — exactly this kind of architecture.

Massive Parallelism

The brain contains approximately 86 billion neurons and roughly 100 trillion synaptic connections, most of which can operate simultaneously and independently. Neuromorphic chips pursue the same massive parallelism — millions of neuron cores running concurrently, rather than a small number of cores operating at high speed in sequence.

Ultra-Low Power Consumption

The three properties above combine to yield ultra-low power consumption. Event-driven operation means near-zero energy use during idle periods, in-memory computing eliminates data-movement energy costs, and massive parallelism reduces the clock frequency required of each individual core.

3. Spiking Neural Networks

The core algorithmic model running on neuromorphic hardware is the spiking neural network (SNN).

In traditional artificial neural networks, neurons transmit continuous values (e.g., the output of a ReLU is a floating-point number). In SNNs:

Neurons communicate through discrete spikes — they either fire a spike or remain silent. Information is encoded in the temporal pattern of spikes.

Key properties of SNNs:

Temporal coding: Information resides not only in "whether a spike occurs" but in "when it occurs." The time interval between two spikes and the synchronous firing patterns of multiple neurons all carry information
Sparse activation: At any given moment, only a small fraction of neurons are active while the vast majority remain silent — consistent with observations of the brain
Biological plausibility: SNNs approximate the behavior of real neurons more closely than traditional ANNs

Comparison with Traditional ANNs

Property	Traditional ANN	Spiking Neural Network (SNN)
Signal type	Continuous floating-point values	Discrete spikes
Temporal dimension	Typically ignored	Core encoding dimension
Activation density	Typically dense	Highly sparse
Energy consumption	Full-network computation per forward pass	Only spike-receiving neurons compute
Hardware affinity	GPU	Neuromorphic chips
Training method	Backpropagation (mature)	Alternative learning rules (under development)

4. Major Hardware Platforms (2025–2026)

Intel Loihi Family

Loihi 2 (2021): 128 neuromorphic cores, ~1 million neurons, supports programmable spike-based learning rules
Hala Point (2024): 1,152 Loihi 2 processors, ~1.15 billion neurons and 128 billion synapses, consuming only hundreds of watts
Loihi 3 (expected 2026): Next generation, further improving scale and energy efficiency

IBM NorthPole

NorthPole (2023) is a digital neuromorphic chip optimized for inference. It fully integrates computation and memory on-chip, eliminates off-chip memory access, and demonstrates extremely high energy efficiency on inference tasks.

BrainChip Akida

BrainChip Akida targets commercial deployment with a focus on edge computing. Akida 1.0 is already in commercial use for edge AI tasks such as vision and audio processing; Akida 2.0 adds enhanced support for Transformer-class models. It demonstrates that neuromorphic computing can make its way into real products.

BrainScaleS-2

BrainScaleS-2 is a product of the European Human Brain Project. It uses analog computation and contains 512 adaptive leaky integrate-and-fire neurons. It operates on an accelerated time scale — roughly 1,000 times faster than biological real time — making it suitable for research requiring long-duration simulations.

5. Progress in China: SpikingBrain 1.0

In 2025, a Chinese research team released SpikingBrain 1.0 (瞬悉 1.0), a large-scale model based on spiking neural networks. Its key features include:

High training efficiency: Achieves high-performance training with very limited data
Architecture conversion: Supports converting existing Transformer models into spiking architectures, fully leveraging pre-trained weights
End-to-end spiking computation: Inference is performed entirely with spike signals, making it compatible with neuromorphic hardware

The "model conversion" approach of SpikingBrain 1.0 is particularly noteworthy. Rather than training an SNN from scratch, it "translates" existing, large-scale pre-trained Transformer models into spiking versions. This sidesteps the bottleneck of difficult SNN training while preserving the energy-efficiency advantages of spiking computation during inference.

6. Performance and Energy Efficiency

The core advantage of neuromorphic computing lies in its energy-efficiency ratio. As a rough estimate:

On certain tasks, neuromorphic chips can achieve approximately 1,000 times the energy efficiency of GPUs.

This figure results from the compounding of multiple factors:

Event-driven operation eliminates idle power consumption
In-memory computing eliminates data-movement power consumption
Spike sparsity reduces the actual volume of computation
Low-precision computation (spikes are inherently 1-bit signals)

However, it is important to note that this advantage manifests primarily during inference and depends on the sparsity of the task. For tasks requiring dense floating-point operations (such as large-scale matrix multiplication), traditional GPUs still hold the upper hand.

7. Connection to Predictive Coding and Active Inference

Neuromorphic hardware has a natural affinity with brain-inspired algorithms.

The hierarchical prediction and error propagation of predictive coding map naturally onto SNNs: prediction signals are transmitted via feedback connections as spikes, errors are encoded through feedforward spikes, and local learning rules (such as STDP) implement weight updates. The message passing and belief updating of active inference are equally well suited to neuromorphic architectures: probabilistic beliefs are encoded as population spike firing rates, and online belief updating aligns with the real-time processing characteristics of the chips.

Spiking neural networks are the natural physical substrate for implementing predictive coding and active inference.

8. Current Limitations

Despite the exciting prospects, neuromorphic computing still faces formidable practical challenges:

Training Difficulties

The discrete nature of spikes in SNNs prevents the direct application of traditional backpropagation — the gradient of the spike function is zero at most points. Current alternatives include:

Surrogate gradients: Approximating the gradient of the spike function with a smooth function
ANN-to-SNN conversion: Training a traditional ANN first, then converting it to an SNN (the strategy used by SpikingBrain 1.0)
Biological learning rules: Such as STDP, though their effectiveness on complex tasks remains limited

Each of these methods has drawbacks, and none currently matches backpropagation in both performance and efficiency.

Immature Programming Models

Programming for neuromorphic chips is fundamentally different from conventional programming. No mature framework exists that plays the role PyTorch or TensorFlow plays for GPUs. Although Intel's Lava framework and BrainChip's MetaTF are making progress, the ecosystem is still in its very early stages.

Weak Ecosystem

Compared to the GPU ecosystem (CUDA, cuDNN, PyTorch, vast collections of pre-trained models), the neuromorphic ecosystem is virtually nonexistent. This means that even if the hardware has a theoretical energy-efficiency advantage, developers lack the tools and models to exploit it.

Scaling Challenges

The largest neuromorphic systems to date (such as Hala Point's 1.15 billion neurons) remain far smaller than the brain's 86 billion neurons. How to scale further while maintaining energy-efficiency advantages is an open engineering problem.

9. The Full-Stack Vision

The ultimate vision of neuromorphic computing extends beyond a single chip to a complete technology stack spanning from theory to hardware:

Theory layer: Free Energy Principle (unified mathematical framework)
    ↓
Algorithm layer: Predictive Coding + Active Inference (brain-inspired algorithms)
    ↓
Model layer: Spiking Neural Networks (brain-inspired computational models)
    ↓
Hardware layer: Neuromorphic Chips (brain-inspired hardware)

Each layer draws inspiration from how the brain works, and the layers are naturally compatible with one another. In today's mainstream AI, every layer of the stack (statistical learning theory, backpropagation, ANNs, GPUs) is highly mature and mutually well-adapted. Every layer of the brain-inspired full stack is still in its early stages, but its potential advantages — energy efficiency, data efficiency, online learning, and a unified perception-action framework — are precisely what current AI lacks most.

10. Summary

Neuromorphic computing is not simply "switching to a different chip." It represents a comprehensive paradigm shift from the lowest hardware levels to the highest theoretical levels — from clock-driven to event-driven, from separated memory and compute to in-memory computing, from continuous activations to discrete spikes, from passive inference to active inference.

The complete logical chain:

Current AI hardware (GPUs) differs fundamentally from the brain's computational approach, with an energy-efficiency gap of roughly three orders of magnitude
Neuromorphic chips mimic brain architecture through event-driven computing, in-memory computing, and massive parallelism
Spiking neural networks are the core computational model for neuromorphic hardware, approximating biological neurons more closely than traditional ANNs
Major platforms include Intel Loihi/Hala Point, IBM NorthPole, BrainChip Akida, and BrainScaleS-2, among others
China's SpikingBrain 1.0 explores a practical pathway for converting Transformer models into spiking architectures
Neuromorphic chips can achieve approximately 1,000 times the energy efficiency of GPUs for inference
SNNs are the natural physical substrate for implementing predictive coding and active inference
The main bottlenecks currently lie in the immaturity of training methods, programming models, and the ecosystem
The ultimate vision is a complete brain-inspired technology stack from the Free Energy Principle down to neuromorphic hardware