Zhejiang University’s “Wukong” Neuromorphic Computer: A New Milestone in Brain-Inspired Computing

On August 2, 2025, Zhejiang University’s National Key Laboratory of Brain-Machine Intelligence made a significant announcement that has captured the attention of researchers and technology enthusiasts worldwide. The laboratory unveiled Darwin Monkey, affectionately named “Wukong” (Chinese for “Monkey King”), the latest generation of neuromorphic computing system that has set a new global benchmark in the field. This isn’t just another incremental improvement in computing technology—it represents a fundamental shift in how we approach artificial intelligence and brain simulation.

What Exactly Is a Neuromorphic Computer?

Before we dive into the specifics of “Wukong,” let’s establish a clear understanding of what a neuromorphic computer actually is. The term “neuromorphic” comes from “neuron” (nerve cell) and “morphic” (form or structure), essentially meaning “brain-shaped” computing.

Unlike traditional computers that follow the von Neumann architecture—where data processing and storage are separate—neuromorphic computers mimic the structure and function of biological brains. They use artificial “neurons” and “synapses” to process information in a way that resembles how our own brains work.

Think about it this way: your laptop or smartphone processes information in a very structured, sequential manner, like a factory assembly line. A neuromorphic computer, by contrast, works more like a bustling city where information flows through multiple interconnected pathways simultaneously. This architecture enables more efficient processing for certain types of tasks, particularly those involving pattern recognition, sensory processing, and adaptive learning.

The “Wukong” Breakthrough: By the Numbers

The Darwin Monkey (“Wukong”) system represents a quantum leap in neuromorphic computing capabilities. Let’s examine what makes this system so remarkable:

• 2+ billion pulse neurons: This is the system’s most impressive specification. To put this in perspective, it approaches the neural scale of a macaque monkey’s brain (which has approximately 2.5 billion neurons).

• Over 100 billion synapses: These represent the connections between artificial neurons, enabling complex information processing.

• 960 self-developed Darwin 3rd generation neuromorphic chips: Each server contains 64 of these specialized chips, with the entire system comprising 15 blade-style neuromorphic servers.

• Approximately 2000 watts power consumption: Remarkably energy-efficient considering its scale—comparable to running ten household vacuum cleaners simultaneously.

• World’s first: It’s the first neuromorphic computer based on specialized neuromorphic chips to exceed 2 billion neurons.

This achievement positions “Wukong” significantly ahead of previous benchmarks. For context, Intel’s Hala Point system, released in April 2024, supported 1.15 billion neurons—meaning “Wukong” has nearly doubled that capacity.

The Evolution: From “Mickey” to “Wukong”

“Wukong” didn’t emerge in isolation. It represents the culmination of years of dedicated research by Zhejiang University’s team. In September 2020, the same research group made headlines with Darwin Mouse, nicknamed “Mickey,” China’s first neuromorphic computer capable of supporting billion-scale neurons.

The progression from “Mickey” to “Wukong” in less than five years demonstrates remarkable acceleration in the field. This isn’t merely about scaling up numbers—it reflects fundamental advances in chip design, system architecture, and software optimization that make such massive neural networks possible.

Inside the Darwin 3rd Generation Chip

At the heart of “Wukong” lies the Darwin 3rd generation neuromorphic computing chip, jointly developed by Zhejiang University and Zhijiang Laboratory in early 2023. Each of these specialized chips is an engineering marvel:

• Supports more than 2.35 million pulse neurons
• Handles billion-level synapses
• Features a dedicated instruction set for neuromorphic computing
• Incorporates neuromorphic online learning mechanisms

What makes these chips particularly impressive is how they translate biological principles into silicon. Traditional computer chips process information using binary signals (0s and 1s) in a clock-driven, synchronous manner. Darwin 3 chips, however, use “spikes” or “pulses” to transmit information—much like biological neurons communicate through electrical impulses.

This pulse-based approach offers significant advantages:

• Energy efficiency: Neurons only consume power when they fire (send a pulse), unlike traditional processors that constantly draw power
• Event-driven processing: The system responds to changes in input rather than processing data at fixed intervals
• Temporal processing: It can naturally handle time-dependent information, which is crucial for sensory processing and real-world interaction

The Engineering Marvel Behind “Wukong”

Building a system with over 2 billion artificial neurons presents extraordinary engineering challenges. The research team spent more than two years overcoming these hurdles, achieving breakthroughs in several critical areas:

1. Large-Scale Neural System Interconnection Architecture

Creating a system where billions of artificial neurons can communicate effectively requires sophisticated interconnection. The team developed a hierarchical, scalable chip-to-chip interconnection system based on a multi-dimensional grid topology. This architecture allows for efficient communication between the 960 Darwin 3 chips while maintaining system stability.

2. Adaptive Time Step Control Method

When coordinating billions of artificial neurons, timing becomes critical. The team implemented an adaptive time step control method that enables large-scale neural coordination. This innovation ensures that different parts of the system can operate in harmony without the synchronization issues that plague conventional parallel computing systems.

3. Wafer-Level Integration: DarwinWafer

Perhaps the most revolutionary advancement is the development of DarwinWafer, a wafer-level super-integrated neuromorphic computing chip. By leveraging domestic wafer substrate technology and CoWoS-S 2.5D advanced packaging technology, the team created a system that transcends traditional printed circuit board (PCB) limitations.

The significance of this achievement cannot be overstated. Traditional server designs connect individual chips on a circuit board, creating bottlenecks in communication speed and energy efficiency. DarwinWafer eliminates these bottlenecks by integrating 64 Darwin 3 chip dies directly onto a single 12-inch wafer.

This “System on Wafer” (SoW) approach represents a paradigm shift in computer architecture:

• Micro-nano scale interconnect optimization: Connections between chips are dramatically shortened
• Elimination of traditional photomask constraints: Enables more flexible design layouts
• Enhanced communication speed: Data can travel between chips much faster
• Improved energy efficiency: Reduced distances mean less power needed for data transfer

Each blade-style neuromorphic server in the “Wukong” system can be constructed using either traditional PCB-based interconnects or this revolutionary wafer-level integration, giving the system remarkable flexibility.

4. Hierarchical Resource Management Framework

Managing resources in a system with billions of artificial neurons requires sophisticated software. The team developed a layered system resource management framework with specialized data swapping strategies for multi-level memory systems. This framework ensures that the massive neural network can be efficiently allocated and managed.

The Darwin Neuromorphic Operating System

Hardware alone isn’t enough—specialized software is needed to unlock the full potential of neuromorphic hardware. The research team developed a new generation of Darwin neuromorphic operating system with several key innovations:

• Layered resource management architecture: Organizes system resources in a hierarchical manner for efficient allocation
• Load-aware scheduling algorithms: Dynamically adjusts resource allocation based on task demands
• Dynamic time slice allocation mechanism: Optimizes task execution timing for maximum efficiency
• Communication bandwidth and task characteristic considerations: Takes into account both communication needs and task properties when scheduling operations

This operating system is crucial because neuromorphic computing presents fundamentally different challenges than traditional computing. Instead of managing processes and threads, it must manage the dynamic connections between billions of artificial neurons and synapses. The Darwin OS bridges this gap, making the complex hardware accessible for practical applications.

Real-World Applications: Beyond Theoretical Promise

What truly sets “Wukong” apart is its practical applicability. The system isn’t just a research curiosity—it’s already demonstrating real value in several domains:

1. Neuromorphic Large Language Models

The team has successfully deployed DeepSeek neuromorphic large models on “Wukong” that can perform logical reasoning, content generation, and mathematical problem-solving. These models leverage the system’s unique architecture to process information in ways that traditional AI models cannot.

2. Biological Brain Simulation

One of the most exciting applications is the simulation of biological brains across various scales:

• Nematodes (roundworms): Simple nervous systems with approximately 302 neurons
• Zebrafish: More complex brains with about 100,000 neurons
• Mice: Mammalian brains with approximately 71 million neurons
• Macaques: Primate brains approaching 2.5 billion neurons

This capability provides neuroscientists with an unprecedented tool for studying brain function without the ethical and practical limitations of biological experiments.

3. Accelerating Neuroscience Research

As a natural platform for brain simulation, “Wukong” offers neuroscientists a powerful tool for exploring brain mechanisms. Researchers can test hypotheses about neural function, observe how information flows through simulated neural networks, and investigate the basis of cognition, memory, and learning—all in a controlled computational environment.

Why “Wukong” Matters: Three Transformative Impacts

The significance of “Wukong” extends far beyond its impressive specifications. It represents a strategic advancement with three major implications:

1. A New Computing Paradigm for Artificial Intelligence

Current AI systems, particularly large language models, face significant challenges:

• High energy consumption: Training large models can consume as much energy as several households use in a year
• Massive computational requirements: Demanding expensive hardware infrastructure
• Limited online learning capabilities: Most models require extensive retraining for new information

“Wukong” addresses these challenges through its neuromorphic architecture:

• Energy efficiency: The event-driven nature of pulse neurons dramatically reduces power needs
• Scalability: The architecture naturally supports massive parallelism
• Online learning: The system can continuously adapt and learn from new information without complete retraining

This could pave the way for AI systems that are not only more powerful but also more sustainable and adaptable.

2. A Powerful Tool for Brain Science

Neuroscience has long struggled with the complexity of the brain. “Wukong” provides a unique platform for brain simulation that could accelerate discoveries in several ways:

• Hypothesis testing: Scientists can test theories about brain function in silico before conducting biological experiments
• Disease modeling: Simulating neurological conditions could lead to better understanding and treatments
• Developmental studies: Observing how neural networks develop and change over time
• Cross-species comparison: Studying similarities and differences between neural architectures across species

This computational approach complements traditional neuroscience methods, potentially reducing the need for certain types of animal experiments while accelerating discovery.

3. A Stepping Stone Toward Artificial General Intelligence

While today’s AI excels at specific tasks, it lacks the flexible reasoning and general intelligence of humans. “Wukong” represents a different approach—one that draws inspiration from the most intelligent system we know: the human brain.

By mimicking the brain’s architecture while surpassing its computational speed, neuromorphic systems like “Wukong” could help bridge the gap between narrow AI and artificial general intelligence (AGI). The system’s ability to handle temporal information, learn continuously, and process sensory data efficiently aligns more closely with human cognition than traditional AI approaches.

Frequently Asked Questions

What makes neuromorphic computing different from traditional AI hardware?

Traditional AI hardware like GPUs and TPUs are optimized for matrix operations used in deep learning. They process information in batches using high-precision numbers. Neuromorphic computing, by contrast, uses spiking neural networks that communicate through discrete events (spikes), similar to biological neurons. This approach is more energy-efficient for certain tasks and better suited for real-time, event-driven processing.

How does “Wukong” compare to the human brain?

While “Wukong” approaches the scale of a macaque monkey’s brain (2.5 billion neurons), it’s still far smaller than the human brain, which contains approximately 86 billion neurons. More importantly, the human brain’s complexity comes not just from neuron count but from intricate connectivity patterns, multiple neuron types, and sophisticated biological mechanisms we don’t fully understand. “Wukong” represents progress toward brain-scale systems, but true brain emulation remains a distant goal.

Why is the power consumption of “Wukong” significant?

At approximately 2000 watts during typical operation, “Wukong” achieves remarkable energy efficiency considering its scale. Traditional supercomputers performing similar-scale neural network simulations would require tens or hundreds of kilowatts. This efficiency stems from the event-driven nature of neuromorphic computing—components only consume power when active, unlike traditional processors that constantly draw power.

What practical applications might emerge from this technology?

Near-term applications include:

• Low-power edge AI devices for real-time sensory processing
• More efficient data center infrastructure for specific AI workloads
• Advanced brain-computer interfaces
• Accelerated neuroscience research
• Robotics with more natural sensory processing and motor control

How does the wafer-level integration (DarwinWafer) improve performance?

Traditional chip packaging creates communication bottlenecks between individual chips. DarwinWafer integrates 64 chip dies directly onto a single wafer, enabling micro-nano scale interconnects. This dramatically reduces communication latency and power consumption between chips while increasing bandwidth—critical factors for coordinating billions of artificial neurons.

What does “pulse neuron” mean in this context?

A pulse neuron (or spiking neuron) is an artificial neuron model that communicates through discrete events (pulses or spikes) rather than continuous values. Information is encoded in the timing and frequency of these pulses, similar to biological neurons. This approach is more biologically plausible and energy-efficient than traditional artificial neurons used in deep learning.

How does “Wukong” support online learning?

The Darwin 3 chips incorporate neuromorphic online learning mechanisms, allowing the system to adapt and learn from new information without complete retraining. This mimics how biological brains continuously update their knowledge based on experience, unlike most traditional AI models that require extensive retraining for new data.

Why is simulating animal brains important?

Simulating biological brains at various complexity levels helps neuroscientists:

• Understand fundamental principles of neural computation
• Study how specific neural architectures support different cognitive functions
• Investigate neurological disorders in controlled environments
• Validate theories about brain function before biological testing
• Develop more biologically inspired AI algorithms

What challenges remain in neuromorphic computing?

Despite “Wukong’s” achievements, several challenges persist:

• Developing programming models and tools for neuromorphic systems
• Creating algorithms that fully leverage neuromorphic hardware advantages
• Scaling to human-brain scale while maintaining stability
• Improving the biological fidelity of neural models
• Integrating neuromorphic systems with traditional computing infrastructure

How might this technology evolve in the next decade?

Future developments might include:

• Larger-scale systems approaching human-brain scale
• More biologically realistic neuron and synapse models
• Hybrid architectures combining neuromorphic and traditional computing
• Specialized neuromorphic chips for specific applications
• Improved software tools and programming environments
• Wider adoption in commercial AI applications

The Road Ahead

“Wukong” represents not an endpoint but a significant milestone on the journey of neuromorphic computing. As research continues, we can anticipate several evolutionary paths:

Scaling Up

The progression from “Mickey” to “Wukong” demonstrates rapid scaling capability. Future systems will likely continue this trajectory, potentially approaching human-brain scale (86 billion neurons) within the next decade. However, scaling isn’t just about neuron count—it’s about creating increasingly sophisticated neural architectures that better mimic biological brains.

Energy Efficiency Improvements

One of neuromorphic computing’s most promising aspects is its potential for extreme energy efficiency. Future systems will likely push this boundary further, enabling powerful AI capabilities in resource-constrained environments like mobile devices and IoT sensors.

Algorithmic Innovation

Hardware advances must be accompanied by new algorithms designed specifically for neuromorphic architectures. As researchers develop a deeper understanding of how to program these systems effectively, we’ll see more sophisticated applications that fully leverage their unique capabilities.

Integration with Traditional Computing

Rather than replacing conventional computing, neuromorphic systems will likely complement them in hybrid architectures. We might see specialized neuromorphic accelerators working alongside traditional processors, handling specific tasks where they excel while conventional systems manage other operations.

A Balanced Perspective

While “Wukong” represents an impressive achievement, it’s important to maintain a realistic perspective. Neuromorphic computing is not a magic solution that will instantly solve all AI challenges. It excels at certain types of problems but may be less suitable for others. The most promising path forward likely involves integrating neuromorphic approaches with other computing paradigms to create more versatile and capable systems.

The true value of “Wukong” lies not just in its specifications but in what it enables: new research directions, novel applications, and a deeper understanding of both artificial and biological intelligence. By providing a platform for exploring brain-inspired computing at unprecedented scale, it opens doors to discoveries we can’t yet fully anticipate.

Conclusion: The Significance of “Wukong” in Context

Zhejiang University’s “Wukong” neuromorphic computer marks a significant milestone in our quest to create computing systems that work more like biological brains. Its achievement in scaling to over 2 billion artificial neurons represents not just a technical triumph but a strategic advancement with far-reaching implications.

What makes “Wukong” particularly noteworthy is how it balances scale with practicality. It’s not merely a theoretical exercise but a functional system already demonstrating value in AI applications and brain simulation. The engineering breakthroughs behind it—particularly the wafer-level integration and sophisticated resource management—solve real problems that have constrained previous neuromorphic systems.

This development also highlights China’s growing leadership in advanced computing research. The progression from “Mickey” to “Wukong” demonstrates sustained commitment and capability in this critical field. As neuromorphic computing continues to evolve, systems like “Wukong” will play an increasingly important role in shaping the future of artificial intelligence and our understanding of the brain.

For those following the trajectory of computing technology, “Wukong” represents an important data point in the ongoing evolution beyond traditional von Neumann architectures. While it won’t replace conventional computers tomorrow, it points toward a future where specialized computing architectures handle specific tasks more efficiently—creating a more diverse and capable computing ecosystem.

As researchers continue to explore the potential of neuromorphic computing, systems like “Wukong” will serve as both practical tools and inspiration for the next generation of brain-inspired technologies. The journey from simulating nematode worms to approaching primate-scale brains demonstrates remarkable progress, and suggests even more exciting developments lie ahead.

The true measure of “Wukong’s” impact won’t be found in its specifications alone, but in the discoveries it enables, the applications it powers, and the new questions it helps us ask about intelligence—both artificial and biological. In this sense, “Wukong” isn’t just a computer; it’s a window into the future of computing and cognition.