Generative Models on Analog Hardware with Dynamics

Yu-Neng Wang, Sara Achour · 2026

Read Generative Models on Analog Hardware with Dynamics by Yu-Neng Wang, Sara Achour (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-27294.txt.

Introduction

Wang and Achour build generative models that run on analog hardware instead of digital chips. Coupled oscillators and analog Ising machines solve differential equations at a sliver of digital energy cost. A mismatch blocks them: modern generative models want flexible software-defined dynamics, while hardware dynamics are fixed by physics. Their Analog Interaction Systems framework defines what such hardware can express. In the 2026 hardware paper, the expressivity gap against neural baselines is measured directly.

Two hardware-friendly tricks narrow the gap: parameters that switch over time, and hidden physical states. A Wasserstein GAN training method fits the models without forcing them to track one set path. Sparse links and low-bit parameters prove necessary for practical builds. Energy per generated image lands near twenty-three microjoules for their chosen design in the reported estimate, about a hundredfold better than digital baselines. Image quality scores beat prior analog generators severalfold.

Catalog record

CREATORYu-Neng Wang, Sara Achour
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE49 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.27294

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Generative Models on Analog Hardware with Dynamics Yu-Neng Wang, Sara Achour Abstract . Analog hardware platforms such as coupled oscillators and Analog Ising Machines naturally solve differential equations at a fraction of the energy cost of digital computation, making them attractive for low-power generative modeling, yet a fundamental mismatch exists: modern generative models assume flexible, software-defined dynamics, whereas analog hardware imposes fixed, physics-determined differential equations with limited approximation capacity. This paper introduces Analog Interaction Systems (AIS), a unified framework for hardware-implementable dynamical systems, and empirically characterizes their expressivity gap relative to neural network baselines. Two hardware-compatible mechanisms are proposed to narrow this gap — time-varying piecewise parameters and hidden physical states — and a Wasserstein GAN training procedure is developed to enable training of these models without requiring them to follow a specific trajectory. We characterize how area and power scale with connection density and precision, showing that sparse connectivity and low-bit-width quantized parameters are necessary for practical implementation, and estimate an energy cost of 23 ​ μ 23~\mu J per generated image for the chosen architecture, representing a 2-orders-of-magnitude improvement over digital baselines. On MNIST and Fashion-MNIST, our oscillator-based AIS achieves FID scores of 27.6 and 80.8, outperforming the best prior hardware-implementable analog generative models by 3 – 4 × 3\text{--}4\times with a 4-bit sparse architecture. 1. Introduction (a) Conventional. (b) Analog Hardware. Figure 1. Comparison between conventional and analog hardware image generation. The conventional method runs discrete time steps to emulate a differential equation whose vector field is computed by a neural network. The analog hardware executes a continuous differential equation whose vector field is governed by physical interactions. Analog computing offers a fundamentally different computational substrate. Rather than executing discrete arithmetic operations, analog systems allow physical state to evolve continuously under the natural laws governing their substrate—be it the phase dynamics of coupled oscillators ( 8 ) , the amplitude relaxation of optical parametric amplifiers ( 13 ) , or the spin settling of Ising machines ( 19 ) . Computation is performed by physics rather than by transistor switching, enabling orders-of-magnitude improvements in energy efficiency for suitable workloads. In particular, analog hardware based on interactions between physical elements, such as couplings between oscillators, has emerged as a widely studied class of dynamics-based analog hardware platforms that leverage time-evolving behaviors to solve combinatorial optimization and machine learning problems ( 3 ; 1 ; 18 ; 10 ; 35 ; 32 ) . At the same time, modern generative models, such as diffusion models, incur substantial energy costs and have become a target workload for both researchers and companies investing in energy-efficient hardware. State-of-the-art image and audio generation systems based on diffusion and flow matching ( 12 ; 17 ) require iterative numerical integration of learned differential equations at inference time, consuming tens to hundreds of watts per query on modern GPU hardware. Analog hardware naturally solves differential equation systems with device physics at a fraction of the energy cost of digital hardware and has recently drawn attention for low-power machine learning tasks. Generative models are also potentially amenable to analog computation, as they are described by systems of differential equations. Can we directly leverage the intrinsic, low-energy continuous dynamics of analog hardware to implement an energy-efficient differential equation-based generative model? 1.1. Challenges with Analog Generative Models There is a mismatch between the generative model workload and dynamics supported by analog hardware: generative models have dynamics dictated by neural networks and analog hardware implements physics-based differential equation systems based in the physical behaviors of the system (Figure 1 ). These physics-based differential equation-solving hardware have several features that render them difficult to leverage for machine learning tasks: • Non-standard Dynamics. Because the differential equation structure is set for the physical system, it may not be good at approximating all functions. In contrast, kernels implementing neural networks are known to be good at approximating a range of functions ( 9 ; 24 ; 16 ) . • Fixed Complexity. The complexity of the differential equation is fixed and dictated by the degree of the interactions. In contrast, the complexity of neural networks can be increased by adding layers. • Physical Limitations. The system is subject to hardware constraints. Sparser interactions are easier to implement in hardware, parameter precision is limited, and the computation is subject to noise. The focus of this work is as follows: Can we come up with a generative model architecture amenable to implementation on analog hardware that overcomes these limitations? 1.2. Analog Interaction Systems as Generative Models We introduce an Analog Interaction System (AIS), a unified model for hardware-implementable analog dynamical systems whose dynamics are fixed by device physics, and investigate their viability as generative models, developing techniques for overcoming the identified challenges. Expressivity. The functional forms available to analog hardware are severely restricted by device physics. For example, oscillator networks are governed by sinusoidal coupling; Analog Ising Machines use fixed polynomial or sigmoidal nonlinearities. We empirically characterize the expressivity gap between neural network-based dynamical systems and several AIS model classes and identify two hardware-friendly mechanisms to close this gap: • Time-piecewise parameters. We find the complexity of AIS behavior can be effectively increased by using time-piecewise parameters, and provide candidate analog model architectures that can implement this behavior. • Hidden physical elements. We find the complexity of the AIS behavior can be effectively increased by introducing hidden physical elements that interact with the physical elements used to generate the image. We also observe that exact trajectory supervision commonly used in generative model training algorithms, such as flow matching, is infeasible for AIS due to its limited flexibility. We instead train AIS with a Wasserstein GAN that penalizes only the terminal visible state, freeing the physics to discover its own path from noise to data. This training method also supports optimizing over hidden physical elements, which lack a ground-truth trajectory to follow. On MNIST and Fashion-MNIST, our oscillator-based AIS achieves FID scores of 27.6 and 80.8, outperforming the best prior hardware-implementable analog generative models by 3\text{--}4\times ( 35 ; 32 ) . Efficiency. We describe the hardware constraints gating the design of such systems and present a sparse AIS analog architecture that organizes physical elements into a grid, allowing for neighboring interactions, and provide scaling laws for the power consumption of the system with increasing connectivity and parameter bit width. This architecture also has several options for implementing the time-piecewise parameters in hardware. An analog core implemented with oscillators is estimated to consume approximately 23\,\mu\text{J} per generated image, a 2-orders-of-magnitude improvement over digital alternatives. Organization. The remainder of the paper is organized as follows. Section 2 defines the analog interaction system. Section 3 formulates the generative modeling problem and presents the expressivity study, including the hidden-state and time-varying-weight experiments. Section 4 describes the hardware architecture and GAN training procedure. Section 5 evaluates generative quality against digital and analog baselines. Section 6 discusses related work, and Section 7 concludes. Model Dynamics |\theta| Physical substrate Ref. NN MLP \dot{x}=W^{(L)}\sigma(\cdots\sigma(W^{(1)}x+b^{(1)})\cdots)+b^{(L)} \sum_{i=1}^{L}(n_{i}^{2}+n_{i}) Standard neural ODE — AIS Kuramoto \dot{\theta}_{i}=\omega_{i}+\sum_{j}K_{ij}\sin(\theta_{j}-\theta_{i}) n+n(n-1) Oscillators ( 8 ) Kuramoto+SHIL \dot{s}_{i}={-}\lambda_{i,0}\sin(2s_{i})+\lambda_{i,1}\sin(s_{i})+\sum_{j}C_{ij}\sin(s_{i}{-}s_{j}) 2n+n\cdot(n-1) Injection-locked oscillators ( 31 ) PolyNet \dot{x}_{i}=k_{i}x_{i}-x_{i}^{3}+\sum_{j}J_{ij}x_{j} n+n^{2} Analog Ising Machine (Poly) ( 2 ) TanhNet \dot{x}_{i}=-x_{i}+\tanh\!\bigl(\sum_{j}J_{ij}x_{j}\bigr) n+n^{2} Analog Ising Machine (Tanh)

Read the complete text