During brain surgery, the monitor in front of the operating surgeon processes imagery on a millisecond timescale, drawing out brain activity in real time from electrical signals. In medical navigation and in brain-computer interfaces that let human thought operate external devices, error-free synchronization between reality and machine can be a matter of life and death.
Modern computer systems rest on the benefits of the von Neumann architecture, conceived in the 1940s. This structure physically separates the processor, which handles computation, from the memory, which holds data, and shuttles data back and forth between the two. This design gave humanity the versatility to execute any calculation simply by rewriting a program.
More than half a century later, this structure has become a bottleneck to further performance gains. While processor speeds have improved dramatically, the speed of reading data from memory and the bandwidth of data transfer have failed to keep pace. Today, most of the power and time a computer consumes is spent not on computation itself, but on moving data through wires.
The Brain: The Ultimate Parallel-Processing Organ
The delay caused by this back-and-forth data movement presents its harshest challenge in the simulation of human brain activity. Inside the brain, approximately 86 billion neurons and more than 100 trillion synapses operate simultaneously in parallel. The brain does not separate computation and memory into different locations—it performs both within the very same physical space: the synapse.
To reproduce brain activity on a computer, one must continuously solve complex differential equations—describing, for example, changes in a neuron's membrane potential—by breaking them into countless minute time steps. Even when this computation is performed using graphics processing units running in cutting-edge data centers, an enormous number of intermediate variables must constantly be shuttled back and forth between memory and processor.
In digital computation algorithms such as the Runge-Kutta method, advancing a single time step requires multiple memory accesses. For surgical navigation to stay synchronized with the real-world movement of the brain, a response speed of at least under 10 milliseconds is required. With conventional digital architectures, the physical time needed to move data becomes the limiting factor, making it extremely difficult to achieve this speed threshold.
A Natural Phenomenon Once Called the Memory Device's Chronic Affliction
Conductance Drift, Which Throws Memory Into Disarray
To eliminate data movement, in-memory computing technology—which performs computation and memory storage within the same device—has been the subject of research. A leading contender that has been under development is the phase-change memristor. This device applies heat to special alloys such as chalcogenide glass, physically altering the structure of the material.
The device toggles between a crystalline state, where atoms are arranged in an orderly fashion, and an amorphous state, where they are scattered irregularly, and uses the resulting change in electrical resistance—which depends on the ratio between these states—to record information. This mechanism, which exploits phase transitions driven by Joule heating, has drawn great expectations as a next-generation storage medium whose data does not vanish even when the power is switched off.
However, treating phase-change materials as computing elements presents an obstacle that has long plagued researchers. The glassy structure of the amorphous state exists in a thermodynamically unstable condition. As time passes, atoms within the material slowly migrate, spontaneously settling toward a more stable, lower-energy state.
Thermodynamic Instability Brought About by Structural Relaxation
Accompanying this physical process, known as structural relaxation, the device's electrical resistance gradually increases over time. This phenomenon, called conductance drift, indicates that data recorded in the memory device changes on its own as time passes.
The network of atoms frozen into a glassy state by a sharp temperature change continues to undergo minute rearrangements even at room temperature. In a digital storage medium, having a recorded value change by the time it is read out is a fatal flaw. If information cannot be stably retained over the long term, the reliability that forms the foundation of computing collapses.
In the fields of materials engineering and circuit design, how to suppress this spontaneously occurring drift phenomenon and keep resistance values constant has been treated as a major research challenge. Conductance drift was long regarded as the single biggest obstacle preventing the practical application of phase-change memristors—a defect to be eliminated.
Synchronizing Mathematics With Physical Phenomena
Turning a Defect Into a Driving Force
In a paper published in the journal Science, a joint research team led by Professor Yang Yuchao (楊玉超) of Peking University and researcher Song Zhitang (宋志棠) of the Chinese Academy of Sciences fundamentally overturned this view of the defect. Rather than eliminating the physical phenomenon by which a device's resistance value naturally changes over time, they built a circuit design that harnesses it as the driving force propelling the computation forward.
When simulating physical changes in the real world or neural activity in the brain, a computer repeatedly predicts a slightly future state from the current state and its rate of change. Algorithms for neural dynamical systems frequently involve dynamically adjusting the width of the integration step depending on the environment.
The research team proved that the process by which a phase-change material naturally undergoes structural relaxation—its resistance value changing over time—is structurally aligned with the mathematical process of solving the time evolution of a differential equation. They discovered that resistance drift, once regarded as a defect, is actually a natural computing engine capable of translating complex mathematical equations directly into physical law.
Mapping Differential Equations Onto Physical Law
A precisely controlled voltage pulse is applied externally to the device on the chip to set the initial state of the computation. After that, the computation is entrusted to the process by which the phase-change material spontaneously changes its state, following the material's own natural physical laws. The very process by which conductance continuously decreases over time physically embodies an adaptive step-size search algorithm.
After a specified amount of time has elapsed, reading out the resulting electrical resistance value directly yields the solution to the differential equation. Whereas digital circuits repeat logical operations in step with a clock signal, this new chip reads out the physical evolution of matter itself as the computational result.
In a commentary published in Science, researchers at Germany's Forschungszentrum Jülich described this approach as being like processing extracted crude oil directly at the oilfield itself, rather than transporting it to a distant factory. By completely eliminating the step of physically moving data within the circuit, the root cause of computational latency has been removed at its foundation.
Quantitative Performance Achieved by a Mature Process
Breaking the 2.12-Millisecond Barrier in Computational Latency
This architecture, which directly links physical phenomena to computation, demonstrated performance surpassing the limits of conventional digital hardware. The chip developed by the research team packs the entire array responsible for in-memory computation and step-size adjustment into an area of a mere 0.28 square millimeters—about the size of a pinhead.
The latency for a single computational step executed within this space was recorded at 2.12 milliseconds—far below the 10-millisecond benchmark required to track brain movement in real time. Even though the chip's internal clock frequency is 50 megahertz, it drives computation efficiently using 9-stage pipeline processing.
On a high-fidelity cortical surface reconstruction task, the developed chip's performance was compared against hardware at the pinnacle of modern computing resources. The shape of the cerebral cortex is extremely complex, and computing its deformation field three-dimensionally in real time is an extremely demanding task.
Outperforming the NVIDIA A100 on Cerebral Cortex Tasks
The developed chip went head-to-head with state-of-the-art hardware running in data centers and demonstrated a clear advantage.
| Metric | NVIDIA A100 GPU | State-of-the-art ASIC accelerator | Developed phase-change memristor chip |
|---|---|---|---|
| Architecture | Separate computation/memory | Dedicated circuit | In-memory computing (driven by physical phenomena) |
| Manufacturing process | 7 nanometers | Various | 40 nanometers |
| Computation method | Digital | Digital | Analog (natural structural relaxation) |
| Processing speed (brain modeling) | Baseline (1x) | Above GPU | Up to 478.18x GPU |
| Power consumption | Baseline (1x) | Comparison target | Up to 1/24.73 of ASIC |
As shown in the table, compared with an NVIDIA GPU, the chip achieved computational speeds up to 478.18 times faster. Even compared against a state-of-the-art ASIC accelerator optimized for specific applications, eliminating the data movement that consumes so much power reduced power consumption to between 1/11.75 and as much as 1/24.73.
The Disruptive Power of a Mature 40-Nanometer Process
A fact just as important as the performance gains lies in the chip's manufacturing process. This chip was fabricated using the widely available, mature 40-nanometer process.
In recent years, improving processor performance has required using special cutting-edge equipment such as extreme ultraviolet lithography to miniaturize transistors down to just a few nanometers. The research team boosted performance not by relying on transistor miniaturization, but through a fundamental rethinking of the computing architecture.
Built on a technology base that can be mass-produced in existing semiconductor fabs, they implemented computational performance on the extremely demanding task of cerebral cortex simulation that significantly exceeds that of cutting-edge digital processors.
The Scaling Wall Facing Analog Computation
Digital Brain Twins Open the Door to Medical Applications
By removing the fundamental constraint of data movement, new applications in medicine and brain science are coming into reach. Digital brain twin technology, which mimics a patient's brain activity on a computer without delay, directly supports the verification of new drug efficacy and real-time navigation during surgery.
In technology that directly reads brain waves to operate external machines, delays in communication and computation translate directly into patient discomfort or device malfunction. The low latency of the developed chip provides a solid foundation for synchronizing human thought and machine action without error.
Controlling Noise and Physical Variation
On the other hand, challenges unique to analog computation—which relies on physical phenomena in materials—remain. Whereas digital computers reliably retain information based on the presence or absence of electric charge, analog computation depends on continuous physical quantities.
It is difficult to completely eliminate the influence of noise contaminating computational results due to minute manufacturing variations in phase-change materials or changes in ambient temperature. Since the thermodynamic behavior of the material is mapped onto a mathematical model, there is a risk that if an unexpected temperature gradient occurs, the accuracy of the solution could deteriorate significantly.
Verification Challenges for Scaling to Millions of Elements
To simulate even larger and more complex regions, such as the entire brain, the number of devices on the chip must be scaled up to the millions. Developing control methods that maintain overall computational accuracy at such scale is a challenge for the future.
Given the reliance on physical phenomena, it will also be unavoidable to verify how the degradation of phase-change materials over years of long-term operation affects the reliability of computation. Whether this new architecture—which transforms a defect into a computing engine—takes root as a genuine foundational technology depends on proving its long-term stability in real-world environments.