arxiv: 2605.14331 · v1 · submitted 2026-05-14 · 📡 eess.SP · cs.AI· cs.ET· cs.IT· cs.LG· math.IT

Recognition: no theorem link

Analog RF Computing: A New Paradigm for Energy-Efficient Edge AI Over MU-MIMO Systems

Wentao Yu , Vincent W.S. Wong

Authors on Pith no claims yet

Pith reviewed 2026-05-15 02:17 UTC · model grok-4.3

classification 📡 eess.SP cs.AIcs.ETcs.ITcs.LGmath.IT

keywords analog RF computingedge AIMU-MIMOenergy-efficient inferencematrix-vector multiplicationmixed-precisionpassive mixerwireless physical layer

0 comments

The pith

In MU-MIMO systems a base station broadcasts weight-encoded RF waveforms so clients perform neural-network matrix-vector multiplications with passive mixers, cutting client energy use by nearly two orders of magnitude.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows how analog RF computing moves the heavy matrix-vector multiplications of edge inference out of digital processors and into the wireless channel itself. A base station encodes neural-network weights into downlink RF waveforms and transmits them to multiple users; each client multiplies the received waveform by its own input-encoded local signal inside an existing passive mixer. Tractable accuracy and energy models let the system jointly tune base-station beamforming and per-client scaling to meet accuracy targets while respecting power and hardware limits. Under 3GPP channel conditions the resulting design delivers client-side energy reductions approaching 100 times versus conventional digital inference, and mixed-precision inference further lowers that cost.

Core claim

By encoding neural-network weights at the base station and broadcasting them as RF waveforms, clients reuse passive mixers to compute the matrix-vector multiplications that dominate inference energy, achieving ultra-low power operation when base-station beamforming and client scaling are jointly optimized for accuracy, transmit power, and hardware constraints.

What carries the argument

Joint base-station beamforming and client-side scaling optimization that enforces per-layer and per-client accuracy targets while minimizing energy under transmit-power and hardware limits for both uniform- and mixed-precision inference.

Load-bearing premise

The derived tractable models for analog matrix-vector-multiplication accuracy and energy consumption faithfully capture real passive-mixer behavior, wireless-channel effects, and hardware impairments.

What would settle it

A hardware testbed measurement, under realistic 3GPP channel conditions, that compares actual client energy draw and inference accuracy against the paper's predicted two-order-of-magnitude savings.

Figures

Figures reproduced from arXiv: 2605.14331 by Vincent W.S. Wong, Wentao Yu.

**Figure 2.** Figure 2: An illustration of the baseband waveform construction and subcarrier [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: IF-port output power scaling versus LO- and RF-port input powers [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 6.** Figure 6: Comparison of uniform- and mixed-precision inference. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗

**Figure 5.** Figure 5: Energy consumption of analog RF computing-based edge inference. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

read the original abstract

Modern edge devices increasingly rely on neural networks for intelligent applications. However, conventional digital computing-based edge inference requires substantial memory and energy consumption. In analog radio frequency (RF) computing, a base station (BS) encodes the weights of the neural networks and broadcasts the RF waveforms to the clients. Each client reuses its passive mixer to multiply the received weight-encoded waveform with a locally generated input-encoded waveform. This enables wireless receivers to perform the matrix-vector multiplications (MVMs) that account for most of the computation burden in edge inference with ultra-low energy consumption. Unlike conventional downlink transmissions which are optimized for communications, analog RF computing requires a computing-centric physical layer that controls both the analog MVM accuracy and the energy consumption for inference. Motivated by this, in this paper, we propose a physical layer design framework for analog RF computing in MU-MIMO wireless systems. We derive tractable models for computing accuracy and energy consumption for inference, formulate a joint BS beamforming and client-side scaling problem subject to computing accuracy, transmit power, and hardware constraints, and develop a low-complexity algorithm to solve the non-convex problem. The proposed design provides client- and layer-specific accuracy control for both uniform- and mixed-precision inference. Simulations under 3GPP specifications show that analog RF computing can significantly reduce client-side energy consumption by nearly two orders of magnitude compared to digital computing, while mixed-precision inference requires even lower energy consumption than uniform-precision inference. Overall, these results establish analog RF computing over wireless networks as a promising paradigm for energy-efficient edge inference.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a joint beamforming and scaling framework for analog RF neural inference in MU-MIMO that claims large client energy cuts, but those cuts rest on linearized mixer models that may not survive real hardware impairments.

read the letter

The core contribution is a physical-layer design that lets a base station encode neural net weights into RF waveforms so clients can reuse their passive mixers for matrix-vector multiplies. They add per-client and per-layer accuracy targets plus support for mixed-precision inference, then solve the resulting joint beamforming and scaling problem with a low-complexity algorithm. The 3GPP simulations report nearly two orders of magnitude lower client energy than digital baselines, with mixed precision doing even better. That combination of client-specific control and mixed precision is the clearest step beyond earlier RF computing sketches. The tractable accuracy and energy expressions are derived from RF principles and look internally consistent, which is useful for optimization work. The main limitation is that those expressions use linearized mixer models and additive noise. Real passive mixers introduce conversion-loss variation, LO leakage, I/Q imbalance, and frequency-dependent effects that the paper does not fully model. Under the same 3GPP channels those extra terms could tighten the accuracy constraints and reduce the reported energy advantage. The simulations therefore test an optimistic model rather than measured hardware behavior. Readers working on wireless edge AI or analog computing will find the formulation and mixed-precision extension worth reading. The work is coherent on its own terms and shows clear engagement with the problem, so it deserves a serious referee even if the hardware validation needs strengthening.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes analog RF computing as a paradigm for energy-efficient edge AI inference over MU-MIMO systems. The base station encodes neural-network weights into broadcast RF waveforms; each client reuses its passive mixer to perform matrix-vector multiplications by multiplying the received waveform with a locally generated input waveform. Tractable closed-form models are derived for analog MVM accuracy (Section III) and energy consumption (Section IV). These models are used to formulate a joint BS beamforming and client-side scaling optimization problem subject to per-client accuracy, transmit-power, and hardware constraints (Section V), which is solved by a low-complexity iterative algorithm. 3GPP-compliant simulations report nearly two orders of magnitude reduction in client-side energy relative to digital baselines, with further gains from mixed-precision inference.

Significance. If the tractable models prove faithful to hardware, the work would establish a practical route to offload the dominant MVM operations of neural inference onto existing RF front-ends, yielding order-of-magnitude client energy savings while retaining client- and layer-specific accuracy control. The low-complexity algorithm and explicit treatment of mixed-precision inference are concrete strengths that could influence both theory and system design in energy-constrained edge AI.

major comments (3)

[Section III] Section III: The tractable accuracy model for analog MVM is obtained under linearized passive-mixer assumptions and additive noise. It is not shown whether the model remains accurate once conversion-loss variation, LO leakage, I/Q imbalance, and frequency-dependent phase noise—standard impairments in 3GPP channels—are included. Because the subsequent optimization (Section V) enforces accuracy constraints derived from this model, any unmodeled degradation would tighten the feasible set and reduce the reported energy gains.
[Section IV] Section IV: The closed-form energy-consumption expressions are constructed directly from the accuracy model of Section III. Without circuit-level validation or Monte-Carlo simulations that inject measured mixer nonlinearities, it is unclear whether the claimed energy figures (and the two-order-of-magnitude advantage) survive realistic hardware impairments.
[Section V] Section V, Eq. (formulation of joint problem): The beamforming and scaling optimization inherits all modeling assumptions from Sections III–IV. A sensitivity study that perturbs the accuracy expression with the omitted impairments and re-solves the problem would be required to confirm that the headline energy reductions remain intact under more realistic conditions.

minor comments (2)

[Section II] Notation for the per-layer precision parameters and the client-specific scaling factors should be introduced once in Section II and used consistently thereafter to avoid ambiguity in the optimization formulation.
[Simulation section] Figure captions for the simulation results should explicitly state the exact energy-reduction factor (rather than “nearly two orders of magnitude”) and the precise 3GPP channel model parameters used.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment point by point below, indicating the revisions that will be incorporated into the next version of the manuscript.

read point-by-point responses

Referee: [Section III] Section III: The tractable accuracy model for analog MVM is obtained under linearized passive-mixer assumptions and additive noise. It is not shown whether the model remains accurate once conversion-loss variation, LO leakage, I/Q imbalance, and frequency-dependent phase noise—standard impairments in 3GPP channels—are included. Because the subsequent optimization (Section V) enforces accuracy constraints derived from this model, any unmodeled degradation would tighten the feasible set and reduce the reported energy gains.

Authors: We agree that the closed-form accuracy model in Section III is derived under linearized passive-mixer assumptions to ensure tractability. In the revised manuscript we will add a dedicated subsection in Section III that analytically bounds the impact of conversion-loss variation, LO leakage, I/Q imbalance, and phase noise, and we will include additional Monte-Carlo simulations under 3GPP channel models that inject these impairments. These additions will quantify any tightening of the accuracy constraints and will be used to adjust the optimization in Section V accordingly. revision: yes
Referee: [Section IV] Section IV: The closed-form energy-consumption expressions are constructed directly from the accuracy model of Section III. Without circuit-level validation or Monte-Carlo simulations that inject measured mixer nonlinearities, it is unclear whether the claimed energy figures (and the two-order-of-magnitude advantage) survive realistic hardware impairments.

Authors: The energy expressions are intentionally tied to the accuracy model to enable the joint optimization. While a full circuit-level validation lies beyond the theoretical scope of this work, the revised manuscript will include Monte-Carlo simulations that incorporate measured nonlinear mixer models and conversion-loss statistics from the literature. These simulations will confirm that the passive-mixer architecture retains its fundamental energy advantage (nearly two orders of magnitude) even when the listed impairments are present. revision: partial
Referee: [Section V] Section V, Eq. (formulation of joint problem): The beamforming and scaling optimization inherits all modeling assumptions from Sections III–IV. A sensitivity study that perturbs the accuracy expression with the omitted impairments and re-solves the problem would be required to confirm that the headline energy reductions remain intact under more realistic conditions.

Authors: We will add a sensitivity study to the revised Section V. The accuracy expression will be perturbed with additional noise and distortion terms that represent the omitted impairments, after which the joint beamforming and scaling problem will be re-solved. The updated results will demonstrate that the reported energy reductions remain within the same order of magnitude, with only modest shrinkage of the feasible set. revision: yes

Circularity Check

0 steps flagged

Derivation chain self-contained; models derived from RF principles, no reductions to inputs by construction

full rationale

The paper derives tractable closed-form models for analog MVM accuracy and energy consumption directly from RF circuit principles and wireless channel models (Sections III and IV), then uses these expressions to formulate and solve a joint beamforming/scaling optimization (Section V) subject to accuracy and power constraints. Simulations under independent 3GPP channel specifications produce the reported energy reductions as outcomes of the optimization, not as fitted parameters or self-referential definitions. No self-citation load-bearing steps, ansatz smuggling, or renaming of known results appear in the derivation; the central claims remain independent of the input assumptions.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework rests on standard wireless channel and hardware models plus the assumption that tractable closed-form expressions for accuracy and energy can be derived; no new free parameters or invented entities are introduced in the abstract.

axioms (1)

domain assumption Tractable models for computing accuracy and energy consumption can be derived from RF and hardware principles.
Invoked to enable the joint optimization problem formulation.

pith-pipeline@v0.9.0 · 5600 in / 1230 out tokens · 54325 ms · 2026-05-15T02:17:26.563206+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages

[1]

Edge artificial intelligence for 6G: Vision, enabling technologies, and applications,

K. B. Letaief, Y . Shi, J. Lu, and J. Lu, “Edge artificial intelligence for 6G: Vision, enabling technologies, and applications,”IEEE J. Sel. Areas Commun., vol. 40, no. 1, pp. 5–36, Jan. 2022

work page 2022
[2]

Computing’s energy problem (and what we can do about it),

M. Horowitz, “Computing’s energy problem (and what we can do about it),” inProc. IEEE Int. Solid-State Circuits Conf. (ISSCC), San Francisco, CA, Feb. 2014

work page 2014
[3]

Deep reinforcement learning for task offloading in mobile edge computing systems,

M. Tang and V . W.S. Wong, “Deep reinforcement learning for task offloading in mobile edge computing systems,”IEEE Trans. Mobile Comput., vol. 21, no. 6, pp. 1985–1997, Jun. 2022

work page 1985
[4]

Joint optimal pricing and task scheduling in mobile cloud computing systems,

H. Shah-Mansouri, V . W.S. Wong, and R. Schober, “Joint optimal pricing and task scheduling in mobile cloud computing systems,”IEEE Trans. Wireless Commun., vol. 16, no. 8, pp. 5218–5232, Aug. 2017

work page 2017
[5]

Razavi,RF Microelectronics, 2nd ed

B. Razavi,RF Microelectronics, 2nd ed. Pearson, 2011

work page 2011
[6]

Disaggregated machine learning via in-physics computing at radio frequency,

Z. Gao, S. K. Vadlamani, K. Sulimany, D. Englund, and T. Chen, “Disaggregated machine learning via in-physics computing at radio frequency,”Science Advances, vol. 12, no. 2, pp. 1–10, Jan. 2026

work page 2026
[7]

Multiple access techniques for intelligent and multifunctional 6G: Tutorial, survey, and outlook,

B. Clerckxet al., “Multiple access techniques for intelligent and multifunctional 6G: Tutorial, survey, and outlook,”Proc. of the IEEE, vol. 112, no. 7, pp. 832–879, Jul. 2024

work page 2024
[8]

Deep learning with coherent nanophotonic circuits,

Y . Shenet al., “Deep learning with coherent nanophotonic circuits,”Nat. Photonics, vol. 11, no. 7, pp. 441–446, Jul. 2017

work page 2017
[9]

In-memory computing with resistive switching devices,

D. Ielmini and H.-S. P. Wong, “In-memory computing with resistive switching devices,”Nat. Electron., vol. 1, no. 6, pp. 333–343, Jun. 2018

work page 2018
[10]

Performing mathematical operations with metamateri- als,

A. Silvaet al., “Performing mathematical operations with metamateri- als,”Science, vol. 343, no. 6167, pp. 160–163, Jan. 2014

work page 2014
[11]

A programmable diffractive deep neural network based on a digital-coding metasurface array,

C. Liuet al., “A programmable diffractive deep neural network based on a digital-coding metasurface array,”Nat. Electron., vol. 5, no. 2, pp. 113–122, Feb. 2022

work page 2022
[12]

Robust analog function computation via wireless multiple-access channels,

M. Goldenbaum and S. Sta ´nczak, “Robust analog function computation via wireless multiple-access channels,”IEEE Trans. Commun., vol. 61, no. 9, pp. 3863–3877, Sept. 2013

work page 2013
[13]

Federated learning via over- the-air computation,

K. Yang, T. Jiang, Y . Shi, and Z. Ding, “Federated learning via over- the-air computation,”IEEE Trans. Wireless Commun., vol. 19, no. 3, pp. 2022–2035, Mar. 2020

work page 2022
[14]

Integrated sensing, communication, and computation over- the-air: MIMO beamforming design,

X. Liet al., “Integrated sensing, communication, and computation over- the-air: MIMO beamforming design,”IEEE Trans. Wireless Commun., vol. 22, no. 8, pp. 5383–5398, Aug. 2023

work page 2023
[15]

Task-oriented over-the-air computation for multi-device edge AI,

D. Wenet al., “Task-oriented over-the-air computation for multi-device edge AI,”IEEE Trans. Wireless Commun., vol. 23, no. 3, pp. 2039–2053, Mar. 2024

work page 2039
[16]

AirNN: Over-the-air computation for neural networks via reconfigurable intelligent surfaces,

S. G. Sanchezet al., “AirNN: Over-the-air computation for neural networks via reconfigurable intelligent surfaces,”IEEE/ACM Trans. Netw., vol. 31, no. 6, pp. 2470–2482, Dec. 2023

work page 2023
[17]

AirFC: Designing fully connected layers for neural networks with wireless signals,

G. Reus-Muns, K. Alemdar, S. G. Sanchez, D. Roy, and K. R. Chowd- hury, “AirFC: Designing fully connected layers for neural networks with wireless signals,” inProc. ACM MobiHoc, Washington, DC, Oct. 2023

work page 2023
[18]

Implementing neural net- works over-the-air via reconfigurable intelligent surfaces,

M. Hua, C. Bian, H. Wu, and D. G ¨und¨uz, “Implementing neural net- works over-the-air via reconfigurable intelligent surfaces,”IEEE Trans. Wireless Commun., vol. 25, pp. 11 562–11 576, Feb. 2026

work page 2026
[19]

Analog computing for signal processing and communications—part I: Computing with microwave networks,

M. Nerini and B. Clerckx, “Analog computing for signal processing and communications—part I: Computing with microwave networks,”IEEE Trans. Signal Process., vol. 73, pp. 5183–5197, Dec. 2025

work page 2025
[20]

Analog computing for signal processing and communications – part II: Toward gigantic MIMO beamforming,

M. Nerini and B. Clerckx, “Analog computing for signal processing and communications – part II: Toward gigantic MIMO beamforming,”IEEE Trans. Signal Process., vol. 73, pp. 5198–5212, Dec. 2025

work page 2025
[21]

MIMO systems aided by microwave linear analog computers: Capacity-achieving architectures with reduced circuit complexity,

M. Nerini and B. Clerckx, “MIMO systems aided by microwave linear analog computers: Capacity-achieving architectures with reduced circuit complexity,”IEEE Trans. Wireless Commun., vol. 25, pp. 14 597–14 610, Mar. 2026

work page 2026
[22]

Efficient processing of deep neural networks: A tutorial and survey,

V . Sze, Y .-H. Chen, T.-J. Yang, and J. S. Emer, “Efficient processing of deep neural networks: A tutorial and survey,”Proc. of the IEEE, vol. 105, no. 12, pp. 2295–2329, Dec. 2017

work page 2017
[23]

NR; user equipment (UE) radio access capabilities (Release 19),

3GPP, “NR; user equipment (UE) radio access capabilities (Release 19),” 3GPP TS 38.306 V19.2.0, Mar. 2026

work page 2026
[24]

ZEM-4300+ Level 7 double balanced mixer,

Mini-Circuits, “ZEM-4300+ Level 7 double balanced mixer,” Datasheet, Rev. C, [Online]. Available: https://www.mouser.ca/datasheet/3/3705/1/ ZEM 4300.pdf

work page
[25]

Majorization-minimization algo- rithms in signal processing, communications, and machine learning,

Y . Sun, P. Babu, and D. P. Palomar, “Majorization-minimization algo- rithms in signal processing, communications, and machine learning,” IEEE Trans. Signal Process., vol. 65, no. 3, pp. 794–816, Feb. 2017

work page 2017
[26]

HAQ: Hardware-aware automated quantization with mixed precision,

K. Wang, Z. Liu, Y . Lin, J. Lin, and S. Han, “HAQ: Hardware-aware automated quantization with mixed precision,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Long Beach, CA, Jun. 2019

work page 2019
[27]

Dynamic precision analog computing for neural net- works,

S. Garget al., “Dynamic precision analog computing for neural net- works,”IEEE J. Sel. Top. Quantum Electron., vol. 29, no. 2, pp. 1–12, Mar. 2023

work page 2023
[28]

Adam: A method for stochastic optimization,

D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” inProc. Int. Conf. Learn. Represent. (ICLR), San Diego, CA, May 2015

work page 2015
[29]

Gradient-based learning applied to document recognition,

Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,”Proc. of the IEEE, vol. 86, no. 11, pp. 2278–2324, Nov. 1998

work page 1998
[30]

5G; study on channel model for frequencies from 0.5 to 100 GHz (Release 19),

3GPP, “5G; study on channel model for frequencies from 0.5 to 100 GHz (Release 19),” 3GPP TR 38.901 V19.3.0, Mar. 2026

work page 2026
[31]

NVIDIA A100 Tensor Core GPU,

NVIDIA, “NVIDIA A100 Tensor Core GPU,” NVIDIA Corporation, Data Sheet, May 2022, [Online]. Available: https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/ a100/pdf/nvidia-a100-datasheet-nvidia-us-2188504-web.pdf

work page 2022
[32]

NR; base station (BS) radio transmission and reception (Release 19),

3GPP, “NR; base station (BS) radio transmission and reception (Release 19),” 3GPP TS 38.104 V19.4.0, Mar. 2026

work page 2026
[33]

NR; user equipment (UE) radio transmission and reception; part 1: Range 1 standalone (Release 19),

3GPP, “NR; user equipment (UE) radio transmission and reception; part 1: Range 1 standalone (Release 19),” 3GPP TS 38.101-1 V19.5.0, Mar. 2026

work page 2026
[34]

CVX: MATLAB software for disciplined convex programming, version 2.2,

CVX Research, “CVX: MATLAB software for disciplined convex programming, version 2.2,” https://cvxr.com/cvx, Jan. 2020

work page 2020