Learning optimal erasure of a Static Random Access Memory
Tomas Basile & Karel Proesmans · 2024
Read Learning optimal erasure of a Static Random Access Memory by Tomas Basile & Karel Proesmans (2024) 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-2411-02044.txt.
Description
Abstract 1 Abstract In this paper, we study the thermodynamic cost associated with erasing a static random access memory. By combining the stochastic thermodynamics framework of electronic circuits with machine learning-based optimization techniques, we show that it is possible to erase an electronic random access memory at arbitrarily fast speed and finite heat dissipation.
Catalog record
| CREATOR | Tomas Basile & Karel Proesmans |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 32 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2411-02044
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Learning optimal erasure of a Static Random Access Memory Tomas Basile & Karel Proesmans Abstract 1 Abstract In this paper, we study the thermodynamic cost associated with erasing a static random access memory. By combining the stochastic thermodynamics framework of electronic circuits with machine learning-based optimization techniques, we show that it is possible to erase an electronic random access memory at arbitrarily fast speed and finite heat dissipation. This disproves a widely held belief that heat dissipation scales linearly with erasure speed. Furthermore, we find driving protocols that minimize the heat dissipation, leading to explicit design principles for future computer memories. This bridges an important gap between the theoretical framework of stochastic thermodynamics and applications in electronic engineering. 2 Introduction The last couple of decades have seen a dramatic increase in the thermodynamic efficiency of computational devices. More specifically, Koomey’s law states that the number of computations, such as erasing a bit, that one can do with a fixed amount of energy has doubled approximately every 1.6 years over the last six decades [ 1 ] . There are however reasons to believe that this scaling will have to come to an end at some point over the next couple of decades. Firstly, Landauer’s limit sets a lower bound on the energy necessary to erase one bit of information, which at current rates would be reached by the end of this century [ 2 ] . Furthermore, there are a number of reasons to believe that Landauer’s limit will never be saturated. Firstly, cache memories, the most energy-consuming memories in modern computers, generally need a constant energy supply, even when not performing any task, which leads to excess heat production. Secondly, a number of finite-time generalizations of Landauer’s principle have been derived that suggest that the energy requirements to erase a bit grows linearly with the speed at which the bit erasure is performed [ 3 , 4 , 5 , 6 ] . If this statement were to be true for real cache memories, it would have profound implications for future computer architectures, as we would reach a fundamental physical limit for standard CMOS technology within the next two decades, where it is physically impossible to further improve computer speed without dissipating more heat. It is therefore important to get quantitative estimates of the fundamental lower limits on the amount of energy needed to operate realistic cache memories, to plan the design and estimate the energy consumption of future CMOS-based computers. The central goal of this paper is to give such an estimate of one of the most important cache memories, namely a static random access memory cells (SRAM cells). To do this, we will use a model that is both thermodynamically consistent, so that one can apply the framework of stochastic thermodynamics to it, and realistic enough to give quantitative predictions on the fundamental minimal thermodynamic cost associated with operating the memory [ 7 , 8 ] . As our focus is on finding the fundamental physical limits, we will not account for typical engineering problems such as dopant fluctuations and limited control. By using a recently-developed algorithm that combines results from stochastic thermodynamics with machine learning [ 9 ] , we find a protocol that minimizes the amount of heat necessary to erase the memory. We find that the total heat production during the erasure process decreases with increasing erasure speed, disproving a widely held belief in the physics community that heat production should increase with increasing erasure speed [ 4 , 5 , 10 , 11 ] . Furthermore, we find that under current design, the minimal amount of heat production is close to the Landauer limit. This suggests that the exponential increase in erasure efficiency could in principle continue for several more decades. In the next section, we will introduce the theoretical background that is necessary to derive the crucial results. In section 4 , we will subsequently present the main results of this paper, namely the optimal protocols and minimal heat production associated with erasing an SRAM cell. These results are then discussed in Section 5 , where we also suggest future research directions. 3 Theoretical background 3.1 Finite-time Landauer principle Landauer’s principle states that there is a minimal thermodynamic cost associated with the processing of information [ 12 ] . More specifically, it states that the amount of heat, Q , produced when erasing a bit, i.e., setting the state of the bit that initially has equal probability to be 0 and 1 to 1 , is bounded by Q\geq k_{B}T\ln 2, (1) where k_{B} is the Boltzmann constant, which we will generally set to 1 , and T is the temperature of the environment. More generally, one has [ 13 ] Q=T\ln 2+T\Delta S, (2) where \Delta S is the total entropy production throughout the process. Eq. ( 2 ) shows that saturating the bound set by Landauer’s principle corresponds to a thermodynamically reversible process, i.e., no net entropy is produced as the entropy increase due to the heat production is exactly compensated by an increase of information on the state of the bit. Such thermodynamic reversibility can only be achieved for arbitrarily slow processes. Finite-time operation generally comes with excess heat production. This has, over the last decade, led to a number of generalizations of Landauer’s principle to finite-time processes [ 3 , 4 , 5 , 6 , 14 , 15 , 16 , 17 , 18 , 19 ] . For bits whose state can be described by a discrete-state Markov systems, one can write a finite-time generalization of Landauer’s principle: Q\geq T\ln 2+\frac{T}{\left\langle\mu\right\rangle t_{f}}, (3) where t_{f} is the duration of the protocol and \left\langle\mu\right\rangle is the expected dynamic activity of the system defined below (cf. Eq. ( 12 )). Similar bounds can also be derived for continuous-state systems [ 3 , 4 ] , quantum mechanical systems [ 6 ] and for erasure with a finite erasure error [ 16 ] . One common thread with all of these systems is that the heat is bounded by a term inversely proportional to the duration of the protocol, due to an increase in entropy production, \Delta S\sim t_{f}^{-1}\Rightarrow Q\sim t_{f}^{-1}, (4) which has also been found in a broad range of toy models and experimental systems, and has been suggested as a universal bound [ 19 , 20 , 21 , 18 , 3 , 4 , 5 , 14 , 22 , 15 , 17 ] . There are however several reasons to question the universality of this scaling law. It seems, at least mathematically, possible to avoid this scaling by letting \left\langle\mu\right\rangle diverge as t_{f}\rightarrow 0 . In fact, there is to the best of our knowledge no fundamental physical law that would forbid such scaling. Furthermore, measurements in real RAM cells show a linear relation between protocol duration and heat production, rather than an inverse proportionality [ 23 ] . This begs the question of whether these memories do not operate at high enough speed yet to observe scaling of the form of Eq. ( 4 ), or whether scaling of this form can actually be omitted by increasing the dynamic activity. 3.2 Thermodynamically consistent