Maximum-Entropy Analog Computing Approaching ExaOPS-per-Watt Energy-efficiency at the RF-Edge

Aswin Undavalli, Kareem Rashed, Zhili Xiao, Arun Natarajan, Shantanu Chakrabartty, Aravind Nagulu · 2025

Read Maximum-Entropy Analog Computing Approaching ExaOPS-per-Watt Energy-efficiency at the RF-Edge by Aswin Undavalli, Kareem Rashed, Zhili Xiao, Arun Natarajan, Shantanu Chakrabartty, Aravind Nagulu (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-2510-24975.txt.

Introduction

Aswin Undavalli, Kareem Rashed, Zhili Xiao, Arun Natarajan, Shantanu Chakrabartty, and Aravind Nagulu build analog compute parts on a max entropy rule to save power. Their frame treats a mesh of analog parts as one phys system whose end state can do math. That end state matches a known spread that can yield dot sums as broad traits of the mesh.

By work in a fast, not yet settled mode, they push speed and power gains past the calm case. A test radio chip, made in a known fine process, shows the plan in real silicon. Reads on that chip point to high compute per watt at the radio edge. The claim rests on that proof chip.

Catalog record

CREATORAswin Undavalli, Kareem Rashed, Zhili Xiao, Arun Natarajan, Shantanu Chakrabartty, Aravind Nagulu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE85 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-24975

Browse the full library · About the Tower of Babel

Full text

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

Maximum-Entropy Analog Computing Approaching ExaOPS-per-Watt Energy-efficiency at the RF-Edge Aswin Undavalli, Kareem Rashed, Zhili Xiao, Arun Natarajan, Shantanu Chakrabartty, Aravind Nagulu Abstract In this paper, we demonstrate how the physics of entropy production, when combined with symmetry constraints, can be used for implementing high-performance and energy-efficient analog computing systems. At the core of the proposed framework is a generalized maximum-entropy principle that can describe the evolution of a mesoscopic physical system formed by an interconnected ensemble of analog elements, including devices that can be readily fabricated on standard integrated circuit technology. We show that the maximum-entropy state of this ensemble corresponds to a margin-propagation (MP) distribution and can be used for computing correlations and inner products as the ensemble’s macroscopic properties. Furthermore, the limits of computational throughput and energy efficiency can be pushed by extending the framework to non-equilibrium or transient operating conditions, which we demonstrate using a proof-of-concept radio-frequency (RF) correlator integrated circuit fabricated in a 22 nm SOI CMOS process. The measured results show a compute efficiency greater than 2 Peta ( 10 15 10^{15} ) Bit Operations per second per Watt (PetaOPS/W) at 8-bit precision and greater than 0.8 Exa ( 10 18 10^{18} ) Bit Operations per second per Watt (ExaOPS/W) at 3-bit precision for RF data sampled at rates greater than 4 GS/s. Using the fabricated prototypes, we also showcase several real-world RF applications at the edge, including spectrum sensing, and code-domain communications. keywords Analog computing, Correlators, Radio-frequency, Compute-in-memory, Edge computing † † equal-contributors: These authors contributed equally to this work. † † equal-contributors: These authors contributed equally to this work. † † equal-contributors: These authors contributed equally to this work. 1 Introduction Advances in artificial intelligence and machine learning have created opportunities for new computing platforms that can trade off computational precision for speed and energy efficiency Horowitz 2014 . This trend has led to the resurgence of analog computing techniques, which exploit primitives such as charge or current conservation and the continuous-time dynamics of the substrate to achieve high throughput and energy efficiency. Unlike digital systems that rely on discrete and clocked logic, analog computing systems operate at the speed of physical processes, enabling inherently parallel and energy-efficient computation Chakrabartty and Cauwenberghs 2007 . In a conventional analog computing architecture, atomic operations like multiplications are implemented using physical multiplicative processes such as Ohm’s law Danial et al. 2019 ; Wan et al. 2022 , translinear principles Kim et al. 2022 ; Chakrabartty and Cauwenberghs 2007 , or capacitor charging Li et al. 2024 . These multiplier units are then physically connected together in a crossbar configuration (shown in Fig. 1 a for two vectors {\bf x} and {\bf y} ) to implement vector operations like inner products Kim et al. 2022 or correlations Li et al. 2024 . While this modular design paradigm forms the core of many state-of-the-art analog machine learning and neuromorphic hardware architectures Chakraborty et al. 2020 , the approach has several scaling limitations. First, computing using these physical multipliers limits the choice to specific devices, such as memristors Danial et al. 2019 , transistors in the sub-threshold or triode regimes Kim et al. 2022 ; Chakrabartty and Cauwenberghs 2007 , and digitally tunable capacitors Li et al. 2024 . Second, the array’s speed, energy-efficiency, and scalability are limited by the interconnect bandwidth, cross-talk between the physical multipliers and the settling-time of the peripheral read-out circuits Amin et al. 2022 . In this paper, we propose an analog computing paradigm where the entire array, including the interconnects, functions as a mesoscopic physical system, utilizing only statistical laws like the physics of entropy production for computation Landauer 1961 ; Peliti and Pigolotti 2021 . The inputs {\bf x} and {\bf y} enforces the boundary conditions of the mesostate ensemble, as depicted in Fig. 1 b, which could be formed by the energy-barriers due to transistors and the strength of the coupling between the energy-barriers is determined by the interconnects and the parasitic capacitances. As shown in Fig. 1 c, each individual mesostate are collections of microstates (for example electron configurations or current paths) and for this work we will assume that the stochastic dynamics of the microstates (as shown in Fig. 1 c) operate at a much faster time-scale than the mesoscopic time-scales Peliti and Pigolotti 2021 . Under the influence of the boundary constraints, the equilibrium or the maximum-entropy configuration of the mesostates encodes the computation of interest. In this work we show that when rectangular symmetry constraints are enforced on the inputs, a macroscopic ensemble potential tracks a monotonic function G({\bf x}^{T}{\bf y}) of the inner-product or the correlation, as shown in Fig. 1 b. While this approach is general enough to be applied to many mesoscopic physical ensembles, in this paper, we demonstrate the proof-of-concept and associated performance benefits using a custom silicon integrated circuit, also shown in Fig. 1 b. Specifically, we leverage the result that a Margin Propagation (MP) computing framework Gu and Chakrabartty 2012 is a maximum-entropy configuration of a mesoscopic ensemble with a rectangular symmetry boundary conditions. The symmetry is illustrated in Fig. 1 c for two ensembles each comprising of four potential wells E^{\pm}_{1-2} and E^{\pm}_{3-4} , whose heights are determined by inputs \{x_{1},y_{1}\} and \{x_{2},y_{2}\} . The rectangular boundary conditions for each ensemble is set according to (x_{1}+y_{1},-x_{1}-y_{1},x_{2}+y_{2},-x_{2}-y_{2}) and (x_{1}-y_{1},-x_{1}+y_{1},x_{2}-y_{2},-x_{2}+y_{2}) . The height of the potential well determines the probability of occupancy of the mesostate and is denoted by (p_{1}^{+},q_{1}^{+},p_{2}^{+},q_{2}^{+}) and (p_{1}^{-},q_{1}^{-},p_{2}^{-},q_{2}^{-}) respectively and shown in Fig. 1 d. Note that the potential wells are shown to be separated in Fig. 1 c, which allows us to specify discrete probabilities for each mesostate, instead of continuous distributions of the underlying microscopic configuration shown in Fig. 1 d. These two ensembles form a part of the larger ensemble as shown in Fig. 1 e, where we have arranged the wells labeled by x_{i}+y_{i},-x_{i}-y_{i} and x_{i}-y_{i},y_{i}-x_{i} on a 2-D grid for the sake of exposition. The corresponding probability for each of the mesostates are specified by p_{i}^{+},q_{i}^{+} and p_{i}^{-},q_{i}^{-} with \sum_{i}^{N}p_{i}^{+}+q_{i}^{+}=1 and \sum_{i}^{N}p_{i}^{-}+q_{i}^{-}=1 respectively. In addition to the rectangular symmetry constraint, we assume an underlying energy-like constraint Tsallis 1988 ; Jaynes 1957 \displaystyle\sum_{i=1}^{N}(x_{i}+y_{i})(p_{i}^{+}-q_{i}^{+})=U^{+}-E_{0}, (1a) \displaystyle\sum_{i=1}^{N}(x_{i}-y_{i})(p_{i}^{-}-q_{i}^{-})=U^{-}-E_{0}, (1b) where U^{+}\geq 0 and U^{-}\geq 0 denote the energy of each ensemble, respectively. If, other than the symmetry and energy constraints, nothing else is assumed about the interactions among the mesostates, then, according to the maximum entropy principle Jaynes 1957 , the most appropriate choice of the steady-state ensemble distribution is the one with the highest entropy. Figure 1: (a) Conventional analog computing architecture for computing inner-products and correlations using multiply-accumulate (MAC) operations. (b) Maximum-entropy analog computing architecture and an integrated circuit prototype for demonstrating the proof-of-concept. Inputs {\bf x},{\bf y} are presented as boundary conditions of a statistical ensemble and the ensemble potential encodes the inner-product/correlation as G({\bf x}^{T}{\bf y}) . (c) Rectangular symmetry constraints on the potential wells with respective heights ( E_{1}^{\pm}=E_{0}\pm x_{1}\pm y_{1} , E_{2}^{\pm}=E_{0}\pm x_{2}\pm y_{2} , E_{3}^{\pm}=E_{0}\pm x_{1}\mp y_{1} , E_{4}^{\pm}=E_{0}\pm x_{2}\mp y_{2} ). (d) Mesostates and microstates corresponding to each potential well and part of a larger ensemble (e). (f) Conceptual illustrations of a system evolving to its maximum-entropy state with

Read the complete text