Energy-Information Trade-Off in Self-Directed Channel Memristors
2025
Read Energy-Information Trade-Off in Self-Directed Channel Memristors (2025) 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-2508-16236.txt.
Introduction
Waleed El-Geresy and co-authors study the energy cost of storing information in a special memory device in a 2025 paper. The device is a self-directed channel memristor, a tiny component whose resistance can hold a value. Such parts are eyed for fresh data storage and brain-like computing, but the energy each stored state takes was poorly mapped.
Lab tests measured the energy needed to set the devices into varied states and how well those states lasted over time. A generative model, built with a conditional adversarial network, then learned the pattern of outcomes for each setting. From that model the team drew energy-information curves for a range of storage delays. The curves show a smooth trade-off: spending more energy buys more usable capacity, and waiting longer costs some of both.
Catalog record
| YEAR | 2025 |
|---|---|
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 22 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-16236
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Energy-Information Trade-Off in Self-Directed Channel Memristors Abstract Understanding the nature of information storage on memristors is vital to enable their use in novel data storage and neuromorphic applications. One key consideration in information storage is the energy cost of storage and what impact the available energy has on the information capacity of the devices. In this paper, we propose and study an energy-information trade-off for a particular kind of memristive device - SDC ( SDC ) memristors. We perform experiments to model the energy required to set the devices into various states, as well as assessing the stability of these states over time. Based on these results, we employ a generative modelling approach, using a cGAN ( cGAN ) to characterise the storage conditional distribution, allowing us to estimate energy-information curves for a range of storage delays, showing the graceful trade-off between energy consumed and the effective capacity of the devices. 1 Introduction and Motivation Memristors are a type of passive electronic device typically defined by their capacity to switch between resistive states. Initially proposed by Leon Chua in 1971 to complete a “periodic table” of passive electronic components [ 1 ] , their potential for describing neuronal dynamics was soon highlighted [ 2 ] . They subsequently emerged as candidate neuromorphic computing elements [ 3 ] , for the storage and in-place processing of information to overcome the Von Neumann bottleneck [ 4 ] . One of the key advantages of neuromorphic computing is energy-efficient information processing. It is therefore essential to develop our understanding of information storage on memristors and how energy available impacts the information capacity of the devices, which plays a key role in enabling their use for intelligent computation. Previous research identified a correlation between the magnitude of current pulses used to program memristive devices and the resulting resistance level achieved [ 5 ] , suggesting a trade-off between power consumption and the final resistance state. In [ 6 ] , an explicit logarithmic trade-off was identified between the final conductance of a device (state) and the energy of the programming pulses applied. Different memristive states can exhibit varying degrees of reliability, with some states being more stable than others due to stochastic resistive drift [ 3 , 7 ] . The instability of these states, coupled with the varying amounts of energy required to program the devices into different states, creates an implicit trade-off between the energy consumed during programming and the information reliably recoverable after different delays. Previous research has attempted to measure the capacity of various memristive devices in the presence of different non-idealities. A graph-theoretic approach to characterizing the capacity of binary memristors in the presence of sneak path current noise was investigated in [ 8 ] . The capacity of PCM ( PCM ) devices for storage was empirically estimated in [ 9 , 10 ] using the Blahut-Arimoto algorithm [ 11 ] . However, these approaches did not consider the energy cost of states. In this paper, we present a study of the information-theoretic limits of information storage under energy constraints on memristive devices in the presence of delay-conditional resistive drift noise. We focus on SDC memristors, which are one of the few types of memristors currently commercially available [ 12 ] . In Section 4 , we use an iterative optimisation procedure to estimate an energy cost function that yields the energy cost of different device initial states. In order to enable us to model the effect of data storage under different delay conditions, in Section 5 , we use a cGAN to model the delay conditional storage distribution. We combine these results in Section 6 , to estimate the cost-constrained capacity for a variety of storage delays. To the best of our knowledge, this is the first study of its kind to explore the explicit trade-off between energy and information capacity in memristors. 2 Problem Formulation Our goal is to estimate a trade-off between the capacity of a memristor for information storage and the energy available to us to program the device. To this end, we must understand the states of the device, which can be used to store information. 2.1 Memristive State and State Instability In [ 13 ] , a proposed memristive state was introduced for Knowm W+ SDC memristors, accounting for the non-linear VI characteristic of the devices. The state is characterised by the relation between the instantaneous voltage, v(t) , and current, i(r,v) : \displaystyle i(r,v(t)) \displaystyle=r\cdot(G_{m}\cdot v(t)+I_{d}(v(t))) (1) \displaystyle I_{d}(v(t)) \displaystyle=\alpha_{1}\cdot(e^{\beta_{1}\cdot v(t)}-1)+\alpha_{2}\cdot(1-e^{-\beta_{2}\cdot v(t)}) (2) where G_{m} , \alpha_{1} , \alpha_{2} , \beta_{1} , and \beta_{2} are parameters that characterise the VI relation for a particular device, and x is the state variable that determines the instantaneous VI relation. I_{d} represents the diode component of the total current. In this paper, we use the model parameters from [ 13 ] , shown in Table 1 . \alpha_{1} \alpha_{2} \beta_{1} \beta_{2} G_{m} 2.73\cdot 10^{-3} 1.313\cdot 10^{-7} 13.92 2.327\cdot 10^{-6} 4.207 Table 1 : Parameters for the SDC memristor state model. However, the identified state is not stable and changes over time, even in the absence of a stimulating voltage. We can represent this mathematically as a conditional probability distribution, where the state after a delay D , denoted \hat{R} , is a random variable conditional on the initial state, denoted R , and the delay, D : \hat{R}\sim P(\hat{R}|R,D) . This formulation lends itself to a communication-theoretic treatment of the storage problem, where the conditional storage distribution can be seen as a delay-dependent communication channel, P(\hat{R}|R,D) , with the initial state, R , denoting the channel input. We can thus characterise the capacity of this storage channel, under given constraints. 2.2 Energy-Constrained Capacity The capacity-cost function [ 11 ] characterises the trade-off between the achievable inverse of the cost-constrained channel capacity (symbols per bit), C , and the maximum tolerable cost (or expense), \beta . We can characterise the capacity of the storage channel for a given delay by finding the capacity achieving input distribution, P_{R}^{*} , which maximises the mutual information between \hat{R} and R , and thus, achieves the capacity of the channel. Thus: \displaystyle C(d)=\max_{P(R)}I_{P(\hat{R}|R,D)}(R;\hat{R}). (3) Note that, usually, the set to which the capacity achieving input distribution belongs, is unconstrained. If, however, we have an average power (energy) budget for programming, we must modify the optimisation to obtain the energy-constrained capacity, denoted by C_{E} , by examining the maximisation problem over the set of valid input distributions - distributions which satisfy the energy budget condition, for a budget B , given as: \displaystyle\mathbb{E}[\mathcal{E}(R)]\leq B (4) where \mathcal{E}(\cdot) is the energy equation, mapping a given input codeword (resistance, or state) \hat{R}=\hat{r} to an energy cost. We can denote the set of input distributions for which the condition B is satisfied as \Gamma_{B} , thus the energy-constrained capacity (for the reduced set of possible input distributions) is given as follows: \displaystyle C(d)=\max_{P(R)\in\Gamma_{B}}I_{P(\hat{R}|R,D)}(R;\hat{R}) (5) 3 Data Acquisition and Processing We perform experiments on the Knowm W+SDC memristor - a tungsten doped device, with a thin-layer structure, including a Ge2Se3 active layer [ 12 ] . We use a single memristor in series with a 100k\Omega resistor, R_{s} . V_{t} is the voltage measured