⚡ electrical engineering

Continuous-time nonlinear closed-loop in-memory computing for high-accuracy massive MIMO detection

This paper presents a continuous-time nonlinear closed-loop in-memory computing architecture that embeds massive MIMO decoding directly into the physical dynamics of memory arrays and operational amplifiers, enabling high-accuracy detection of complex modulation formats through analog physical optimization and mixed-precision iterative refinement.

Original authors: Piergiulio Mannocci, Giacomo Pedretti, Fabian Böhm, Thomas Van Vaerenbergh

Published 2026-07-07

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Piergiulio Mannocci, Giacomo Pedretti, Fabian Böhm, Thomas Van Vaerenbergh

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

The Big Picture: Solving a Giant Puzzle with Physics

Imagine you are trying to solve a massive, complex puzzle where hundreds of pieces are mixed up, and you need to figure out exactly where each one goes. In the world of wireless internet (specifically "Massive MIMO" systems used in 5G and future networks), this is exactly what happens every time your phone connects to a cell tower. The tower sends out many signals at once, and they get jumbled together by the air and noise. The computer has to "un-jumble" them to read your message.

Doing this on a standard computer is like trying to solve that puzzle by moving one piece at a time, checking if it fits, moving it back, and trying again. It takes a lot of energy and time.

This paper introduces a new way to solve this puzzle. Instead of using a digital computer that calculates step-by-step, the authors built a special machine that uses physics itself to do the math. They created a circuit that naturally "flows" toward the correct answer, just like water flowing downhill to find the lowest point in a valley.

The Problem: Too Much Math for Standard Computers

Current wireless systems are getting smarter, using more antennas to talk to more people at once. But this creates a math problem that is too heavy for standard chips.

Digital Computers: They are like a very fast accountant who adds numbers one by one. They are accurate but get tired (use a lot of energy) when the numbers get huge.
Old "In-Memory" Computers: Previous attempts to speed this up used a method called "In-Memory Computing" (IMC). Think of this as a calculator that does all the math at once by lighting up a grid of lights. However, these old calculators were only good at straight lines (linear math). The jumbled wireless signals, however, are curvy and twisted (nonlinear). To solve them, the old calculators had to take many small, slow steps, which defeated the purpose of being fast.

The Solution: A "Self-Correcting" Circuit

The authors built a new type of machine called a Continuous-Time Nonlinear Closed-Loop IMC. Let's break that down with an analogy:

Imagine a marble rolling on a bumpy, curved surface.

The Surface: The shape of the surface represents the math problem (the jumbled signals).
The Goal: The lowest point in the valley is the correct answer.
The Old Way: A digital computer would calculate the slope, take a tiny step, calculate the new slope, take another tiny step, and so on.
The New Way: This new circuit is like the marble itself. You just drop the marble (the signal), and physics makes it roll down the hill. It doesn't need to be told where to go; the shape of the hill (the circuit's design) naturally guides it to the bottom.

Key Features of this Machine:

It's "Closed-Loop": The machine talks to itself. The output feeds back into the input, creating a loop that helps the marble find the true bottom of the valley faster and more accurately.
It's "Nonlinear": The surface isn't just a smooth slide; it has walls and barriers (like a box). The circuit is designed so the marble can't roll off the edge; it stays within the "box" of valid answers. This is crucial for high-quality internet signals.
It's "Continuous-Time": The marble doesn't stop and start. It flows smoothly. This means the solution appears almost instantly, limited only by how fast electricity can move, not by how fast a processor can count.

The Catch: The Marble is a Bit Shaky

There is a problem with using physical machines for math: they aren't perfect. Real-world components (like the "marble" and the "hill") have tiny imperfections. In the paper's experiment, the circuit was built with memory chips that had about 5 bits of precision (a bit like a ruler with only a few markings). Because of this, the "marble" might roll to a spot that is close to the bottom, but not exactly the bottom.

If you just used this machine alone, the internet connection might be a bit fuzzy, especially for very high-quality signals (like 256-QAM, which carries a lot of data).

The Fix: The "Refinement" Team

To fix the imperfections without building a perfect (and expensive) machine, the authors added a Hybrid System.

Think of it like this:

The Fast, Rough Draft: The physical circuit (the marble) does the heavy lifting first. It finds the answer very quickly and very cheaply, but it's slightly off.
The Slow, Perfect Editor: A small, standard digital computer looks at the rough draft. It calculates exactly how far off the answer is (the "residual").
The Correction: The digital computer tells the physical circuit, "You were off by this much. Try again, but start from this new spot."

They repeat this process a few times. The physical circuit does the fast, rough work, and the digital computer does the precise cleanup.

The Result:

Speed: It's still incredibly fast because the physical circuit does the hard part.
Accuracy: It becomes as accurate as a perfect digital computer.
Energy: It uses a fraction of the energy because the digital computer only does a tiny bit of work (just the cleanup).

What They Actually Proved

The authors didn't just write a theory; they built a prototype chip and tested it.

They simulated a system with 16 to 64 antennas (a "Massive MIMO" system).
They showed that the circuit naturally rolls toward the correct answer, minimizing an "energy function" (like the marble finding the valley).
They proved that even with "shaky" hardware (low precision), adding the "refinement" step allowed them to decode very complex signals (256-QAM) with high accuracy.
They showed that this method is much more energy-efficient than current digital methods, especially as the systems get bigger.

Summary

This paper presents a new way to decode wireless signals by turning the math problem into a physical motion problem. Instead of calculating step-by-step, the circuit lets physics do the work, rolling toward the solution like a ball in a bowl. To fix the small errors caused by imperfect hardware, they added a "check-up" step using a digital computer. The result is a system that is both blazingly fast and highly accurate, promising a future where our wireless networks are faster and use less battery power.

Technical Summary: Continuous-time nonlinear closed-loop in-memory computing for high-accuracy massive MIMO detection

Problem Statement
Modern information processing systems face growing computational demands, particularly in wireless communications where Massive Multiple-Input Multiple-Output (MIMO) detection requires solving high-dimensional inverse problems. While Analog In-Memory Computing (IMC) has emerged as a promising approach to overcome the energy and throughput limitations of digital architectures by co-locating memory and computation, existing IMC systems are predominantly limited to linear algebra workloads. These systems typically encode problems in the equilibrium state of a physical system, relying on open-loop or closed-loop architectures that execute discrete-time iterative algorithms (e.g., ADMM, projected-gradient methods) to handle nonlinear optimization. This reliance on discrete-time steps and algorithmic parameters (step size, iteration count) limits the potential for continuous-time physical optimization. Furthermore, the finite precision of analog hardware often restricts the accuracy of nonlinear solvers, making them unsuitable for high-order modulation formats like 256-QAM without significant performance degradation.

Methodology
The authors propose a continuous-time nonlinear closed-loop IMC architecture designed to solve Box-Constrained Zero-Forcing (BCZF) decoding, a nonlinear optimization problem that approaches Maximum Likelihood (ML) accuracy with polynomial complexity.

Continuous-Time Physical Optimization: Unlike previous approaches that use dynamics merely to determine convergence time, this architecture encodes the optimization problem directly within the continuous-time dynamics of a nonlinear feedback network. The circuit consists of resistive memory arrays storing the channel matrix $H$ and supply-limited operational amplifiers (OpAmps). The channel matrix is embedded in the feedback paths, while the saturation of the OpAmps (clipping at supply voltages $\pm V_s$ ) naturally enforces the box constraints required for BCZF.
Energy Minimization Interpretation: The authors derive a compact analytical model showing that the circuit's trajectories correspond to the minimization of an equivalent energy function. The steady-state solution of the circuit satisfies the first-order optimality conditions of this energy function, effectively performing BCZF decoding through physical relaxation rather than sequential arithmetic instructions.
Hybrid Analog-Digital Iterative Refinement: To address the finite precision of analog memory and peripheral circuitry, the authors extend mixed-precision iterative refinement to nonlinear continuous-time optimization. The framework utilizes a low-precision nonlinear physical solver to generate an initial approximate solution. This solution is then progressively corrected using a high-precision (digital or high-precision IMC) residual computation engine. A local constraint-enforcement block ensures that the corrected estimates remain within the constellation box during the refinement process.

Key Contributions

Extension of Closed-Loop IMC: The work extends closed-loop IMC from equilibrium-based linear algebra to trajectory-based nonlinear optimization by directly encoding the target problem within the continuous-time dynamics of a physical system.
Nonlinear BCZF Implementation: The authors demonstrate that the proposed nonlinear feedback network naturally implements BCZF detection through energy minimization, eliminating the need for explicit time discretization or iterative algorithmic control.
Experimental Validation: The computational dynamics and energy-minimization behavior were experimentally validated using a fabricated 90 nm IMC chip. The study utilized a discrete-time emulation methodology to map continuous-time dynamics onto the available hardware, confirming that the system reproduces predicted trajectories despite hardware nonidealities and limited memory precision.
High-Accuracy Hybrid Architecture: The introduction of an iterative refinement framework enables the recovery of high-accuracy solutions from low-precision physical computation, successfully enabling reliable detection of high-order modulation formats (up to 256-QAM).

Results

Convergence and Latency: Simulations indicate that the convergence time of the continuous-time solver scales approximately as $T_{conv} \propto M\sqrt{N}$ (where $M$ is the QAM order and $N$ is the MIMO dimension), which is significantly faster than the $O(N^{2\sim3})$ complexity of digital solvers. The convergence time remains bounded and largely independent of system size and channel noise.
Detection Accuracy: The nonlinear BCZF solver consistently outperforms linear ZF and MMSE detectors across various system sizes and modulation depths. The relative advantage increases with system size.
Impact of Precision: While BCZF is more sensitive to channel quantization than linear detectors, the hybrid iterative refinement approach effectively mitigates this. Experimental emulation and simulations show that after five refinement iterations, the Bit Error Rate (BER) for 256-QAM detection approaches that of ideal high-precision decoding, despite using 5-bit analog memory arrays with 2% variability.
Benchmarking: Projections for a 14 nm CMOS implementation suggest that the architecture offers superior energy efficiency compared to existing low-precision IMC and digital ASICs in the one-shot regime. In the high-precision regime (BER < $10^{-3}$ ), the hybrid architecture achieves competitive throughput and energy efficiency relative to digital and hybrid alternatives, with energy per bit significantly lower than fully digital solutions.

Significance and Claims
The paper claims to establish continuous-time nonlinear dynamics as a new computational primitive for in-memory computing. By shifting the paradigm from equilibrium-based linear algebra to trajectory-based nonlinear optimization, the work demonstrates a pathway toward efficient physical accelerators for high-accuracy wireless communications. The authors assert that their results unify the efficiency of physical optimization with the accuracy of digital correction, overcoming the traditional limitations of analog computing regarding precision. The work suggests that such continuous-time physical optimization could be extended beyond MIMO detection to broader classes of constrained optimization tasks, although the current study focuses specifically on BCZF decoding. The authors note that while the underlying primitives have been validated, a fully integrated dedicated IC is required to fully validate the projected system-level metrics.