⚛️ quantum physics

Highly Versatile FPGA-Implemented Cyber Coherent Ising Machine

This paper presents a highly versatile FPGA-implemented cyber coherent Ising machine that utilizes FP32 continuous values and flexible sequence control to support various algorithms (including CIM, SB, and Jacobi SOR), achieving N=4096 spins on a single chip with calculation speeds over ten times faster than GPUs to enable previously infeasible applications like CDMA multi-user detection and L0 compressed sensing.

Original authors: Toru Aonishi, Tatsuya Nagasawa, Toshiyuki Koizumi, Mastiyage Don Sudeera Hasaranga Gunathilaka, Kazushi Mimura, Masato Okada, Satoshi Kako, Yoshihisa Yamamoto

Published 2026-07-27

📖 9 min read🧠 Deep dive

CC BY 4.0

Original authors: Toru Aonishi, Tatsuya Nagasawa, Toshiyuki Koizumi, Mastiyage Don Sudeera Hasaranga Gunathilaka, Kazushi Mimura, Masato Okada, Satoshi Kako, Yoshihisa Yamamoto

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the world of computing as a massive, bustling city where every building represents a tiny decision. Sometimes, the city needs to solve a giant puzzle: finding the perfect arrangement of traffic lights to stop jams, organizing a hospital's schedule so no one waits, or even figuring out the best way to pack a suitcase. These are "combinatorial optimization" problems, where you have to pick the best combination of choices from a dizzying number of possibilities. The more choices you have, the harder the puzzle gets, often becoming so complex that even the fastest supercomputers in the world get stuck, taking years to find a good answer.

To tackle these impossible puzzles, scientists have been building special "Ising machines." Think of these not as normal computers, but as magical, physics-based engines. Instead of just crunching numbers one by one, they let thousands of tiny switches (called "spins") dance and interact with each other, like a crowd of people trying to find the most comfortable way to sit in a theater. By watching how these spins settle down, the machine finds the best solution to the puzzle. Recently, researchers have been trying to build these machines using light and quantum physics, but it's like trying to build a skyscraper out of glass: it's fragile, expensive, and hard to connect all the pieces together. So, some smart engineers decided to build a "cyber" version—a digital simulation of this magic machine that runs on a special type of computer chip called an FPGA. This allows them to test the physics of the puzzle without needing a lab full of lasers and mirrors.

The paper you're about to read describes a major upgrade to this digital "cyber" machine. The team, led by researchers from the University of Tokyo and others, has built a highly versatile version that can handle much more complex puzzles than before. Previous digital attempts were like using a calculator that only understands "yes" or "no" (binary) or "maybe" (ternary). This new machine, however, speaks the full language of numbers, using precise decimal values (floating-point numbers) to describe how the puzzle pieces interact. This is a big deal because many real-world problems, like decoding signals in a crowded radio room or reconstructing a blurry MRI image, require that level of precision. If you force those problems into a simple "yes/no" box, the answer gets messy and wrong.

The researchers tested their new machine on two very difficult tasks: acting as a super-fast detective for a crowded radio channel (CDMA multi-user detector) and a super-smart image restorer for medical scans (compressed sensing). They found that their new "cyber" machine could solve these problems more than ten times faster than a standard high-end graphics card (GPU) used in modern computers. Even though the machine had to work a bit harder to handle the precise decimal numbers, its speed and flexibility won out. It's like upgrading from a bicycle that can only go on paved roads to a rugged all-terrain vehicle that can go just as fast, but also handle mud, sand, and steep hills that the bicycle couldn't touch.

The Story of the Cyber Ising Machine

The Problem: The "Glass" City
Imagine you are trying to organize a massive party where everyone has to sit next to their friends but avoid their enemies. If you have 10 people, it's easy. If you have 1,000, it's a nightmare. In the world of physics, this is modeled by something called an "Ising model," where tiny magnets (spins) want to align with some neighbors and oppose others. Finding the perfect seating chart is the same as finding the "ground state" (the lowest energy state) of this system.

For a long time, scientists tried to build real machines using light (lasers) or superconducting circuits to solve these problems. These are called "Coherent Ising Machines" (CIMs). They are cool because they use the laws of physics to solve the math instantly. But building them is like trying to build a city out of glass: it's hard to wire up thousands of connections without everything breaking. Because of this, researchers started creating "cyber" versions—software simulations that mimic the physics of the light machines but run on regular computer chips.

The Old Way: The One-Size-Fits-All (But Broken) Tool
Before this paper, there were already some digital Ising machines running on FPGAs (Field-Programmable Gate Arrays). Think of an FPGA as a Lego board that you can reprogram to become any kind of machine. However, the old versions had some serious limitations:

They were too simple: They could only use "binary" (0 or 1) or "ternary" (0, 1, or -1) numbers for the connections between spins. It's like trying to paint a masterpiece using only a black marker and a white eraser. You can't capture the subtle shades of gray needed for real-world problems.
They couldn't handle "Zeeman terms": In physics, a "Zeeman term" is like an external wind blowing on the spins, pushing them in a specific direction. Many real problems (like finding a signal in a noisy radio) need this "wind" to work. The old machines couldn't handle this wind properly.
They were rigid: If you wanted to change the algorithm (the rules of the game), you often had to rebuild the whole machine.

The New Solution: The Swiss Army Knife
The team in this paper built a new FPGA architecture that fixes all these problems. They call it a "Cyber Coherent Ising Machine." Here is what makes it special:

It speaks "Real" Numbers: Instead of just 0s and 1s, this machine uses "single-precision floating-point" (FP32) numbers. This is like upgrading from a black-and-white TV to a 4K color TV. It can handle the precise, messy decimal values needed for real-world engineering problems.
It Handles the "Wind": It can now properly manage Zeeman terms, allowing it to solve problems like the CDMA multi-user detector (finding who is talking in a crowded room) and L0-norm regularization-based compressed sensing (reconstructing a clear image from very few blurry pieces, like an MRI scan).
It's a Chameleon: The machine is built with a "control module" that acts like a remote control. By simply rewriting the code on this remote, the machine can switch between different modes:
- Open-loop CIM: The original, simpler version.
- Closed-loop CIM: A newer, more complex version that uses "chaotic amplitude control" (a fancy way of saying it shakes things up to escape bad solutions).
- Jacobi SOR: A method for solving systems of equations.
- Simulated Bifurcation (SB): Another popular algorithm that it could run if the code is changed.

The Race: Cyber Machine vs. The GPU
To see if their new machine was actually good, the team put it up against a standard high-end graphics card (an NVIDIA Quadro RTX 8000), which is the kind of chip used for gaming and AI. They tested it on two big challenges:

The Radio Detective (CDMA): They tried to figure out which of 4,096 users was sending a message in a noisy channel.
- Result: The FPGA machine was 11 to 30 times faster than the GPU.
- Accuracy: It found the answers just as well as the GPU, and in some cases (using the "closed-loop" mode), it was even better at finding the perfect solution because the chaotic shaking helped it escape dead ends.
The Image Restorer (L0RBCS): They tried to reconstruct a 64x64 pixel MRI image from only 40% of the data.
- Result: The FPGA machine was 12 to 37 times faster than the GPU.
- Accuracy: The images it produced were just as clear, and again, the "closed-loop" version produced slightly better images by getting closer to the theoretical perfect solution.

The Trade-off: Speed vs. Precision
You might wonder, "If it's so fast, why didn't everyone do this before?" The paper explains that there was a trade-off. The previous FPGA machines (like the Simulated Bifurcation ones) were incredibly fast because they used simple binary numbers and had four times more "workers" (parallel processing units) running at once. The new machine uses precise decimal numbers, which takes more logic resources, so it has fewer workers (only 2,048 "MAC PEs" compared to 8,192 in the old binary machines).

Because it has fewer workers, the new machine takes about four times as many clock cycles to finish one step of the calculation compared to the old binary machines. However, because it can solve problems that the binary machines cannot solve at all (due to the need for precise numbers and Zeeman terms), it is a massive leap forward. It's like having a slower car that can drive off-road versus a fast sports car that can only drive on a highway. For the problems that require off-road driving, the slower car is the only one that gets you there.

What This Means
The paper concludes that this new architecture is a "highly versatile" tool. It proves that you don't need a fragile, expensive physical laser machine to get the benefits of Ising physics. You can build a digital version that is:

Versatile: Can switch between different algorithms on the fly.
Precise: Handles real-world decimal numbers and external forces (Zeeman terms).
Fast: Beats standard GPUs by a factor of 10 or more.

The authors suggest that if they can build a cluster of these machines (connecting many FPGAs together), they could make it even faster, potentially solving even bigger puzzles. For now, they have shown that a digital "cyber" machine can handle the messy, complex math of the real world much better than its simpler, binary cousins, all while running at lightning speed.

Technical Summary: Highly Versatile FPGA-Implemented Cyber Coherent Ising Machine

Problem Statement
Combinatorial optimization problems, such as resource allocation and portfolio optimization, are generally NP-hard, requiring exponential time to solve as problem size increases. While Quantum Ising Machines (QIMs) and Coherent Ising Machines (CIMs) offer potential solutions, physical implementations face significant hurdles in achieving dense coupling and large spin counts ( $N$ ) due to wiring constraints and physical limitations. Consequently, researchers have derived classically computable equations from quantum master equations (e.g., Stochastic Differential Equations for CIMs and Simulated Bifurcation models) to be solved on parallel hardware like GPUs and FPGAs.

However, existing large-scale FPGA implementations of these algorithms (specifically Simulated Bifurcation and CIM) suffer from critical limitations:

Discrete Interactions: They typically use binary or ternary values for coupling matrices ( $J$ ) to save memory and logic resources, which is insufficient for real-world signal processing applications requiring continuous real values.
Lack of Zeeman Terms: They often cannot configure Zeeman terms (external fields), which are crucial for applications like compressed sensing and CDMA multi-user detection.
Algorithm Rigidity: The hardware architecture is often fixed to a specific algorithm; changing the model requires a complete redesign.
Limited Scope: They cannot handle both Ising and Quadratic Unconstrained Binary Optimization (QUBO) problems with Zeeman terms simultaneously or perform alternating optimization schemes required for complex problems like $L_0$ -norm regularization-based compressed sensing ( $L_0$ RBCS).

Methodology
The authors developed a "Cyber Coherent Ising Machine" (Cyber CIM) implemented on a single Xilinx ALVEO U250 FPGA. The architecture is designed to be highly versatile, capable of executing three distinct algorithms on the same hardware modules by simply rewriting the sequence control code:

Open-Loop CIM: Solves Ising and QUBO Hamiltonians without amplitude control.
Closed-Loop CIM: Solves Ising and QUBO Hamiltonians with Chaotic Amplitude Control (CAC) to escape local minima.
Jacobi Successive Over-Relaxation (SOR): Solves simultaneous equations for quadratic optimization.

Key Technical Features:

Precision: Unlike previous FPGA systems using 1-bit or fixed-point arithmetic, this system utilizes single-precision floating-point (FP32) for the coupling matrix ( $J$ ), Zeeman terms ( $g$ ), and time evolution calculations.
Versatile Architecture: The core modules include a Calculation Control Top (CTL TOP) that sequences operations, a Local Field Calculator (CAL H) with 2048 parallel Multiply-Accumulate (MAC) processing elements, and a Time Evolution Calculator (CAL CSR).
Memory Management: To store the $4096 \times 4096$ coupling matrix in FP32 within the FPGA's internal memory, the system stores only the upper triangular matrix and uses a multiplexer (J MUX) to reconstruct the symmetric matrix during calculation.
Alternating Optimization: For $L_0$ RBCS, the system alternates between CIM (optimizing the binary support vector $q$ ) and Jacobi SOR (optimizing the continuous source signal $r$ ).

Experimental Evaluation
The system was benchmarked against an NVIDIA Quadro RTX 8000 GPU using two specific applications that require continuous interactions and Zeeman terms:

CDMA Multi-User Detector: An Ising Hamiltonian optimization problem.
$L_0$ RBCS: A QUBO Hamiltonian optimization problem involving sparse signal recovery.

Experiments were conducted on system sizes of $N=1024$ , $N=2048$ , and $N=4096$ .

Results

Versatility and Accuracy: The FPGA successfully implemented both open-loop and closed-loop CIMs, as well as Jacobi SOR, on the same hardware. For CDMA detection, the FPGA and GPU produced identical solution trajectories for open-loop CIM. For closed-loop CIM, slight trajectory differences were observed, attributed to the FPGA's deviation from strict IEEE 754 rounding rules, which affected the chaotic dynamics of the CAC. However, the closed-loop approach generally yielded lower bit-error rates than open-loop approaches on both platforms.
Reconstruction Performance: In $L_0$ RBCS tasks (both random signals and MRI data), the FPGA and GPU achieved comparable Root Mean Square Error (RMSE) results, with closed-loop CS generally outperforming open-loop CS by finding solutions closer to the theoretical ground state.
Speed: The FPGA implementation demonstrated a significant speed advantage over the GPU.
- For CDMA detection, the FPGA was 13 to 30 times faster than the GPU, depending on the system size and algorithm.
- For $L_0$ RBCS, the FPGA was 12 to 37 times faster than the GPU.
- The authors note that while the FPGA's parallelism (64 MAC units) is one-fourth that of a previous single-FPGA Simulated Bifurcation (SB) implementation (which used 256 MAC units but binary interactions), the use of FP32 allowed for the solution of problems previously intractable on FPGAs.

Significance and Claims
The paper claims that this Cyber CIM represents a significant step forward in FPGA-based Ising machine implementations by prioritizing versatility and precision over raw parallelism density.

Overcoming Limitations: It is the first large-scale FPGA system to support continuous-valued (FP32) interactions and configurable Zeeman terms, enabling applications like CDMA multi-user detection and compressed sensing that were not possible with earlier binary-interaction FPGA systems.
Algorithm Agnosticism: The architecture allows for the execution of multiple algorithms (Open-loop CIM, Closed-loop CIM, Jacobi SOR, and potentially SB) on the same hardware without redesign, adapting to the frequent updates in CIM research models.
Performance: Despite having lower parallelism than binary-interaction FPGA systems, the Cyber CIM achieves calculation speeds more than ten times faster than GPU implementations for complex, real-world optimization problems.
Future Potential: The authors suggest that calculation speeds can be further improved by increasing parallelism through FPGA clustering or by utilizing FPGAs with larger internal memory to eliminate the J MUX bottleneck, potentially allowing for higher clock frequencies.

The Story of the Cyber Ising Machine

More like this