PaperlayerView exact revision on arXiv

Source-linked AI summary

Analog RF Computing for Integrated Data and AI Flows Over Wireless Networks

Wentao Yu, Vincent W. S. Wong

arXiv:2610.20841v1cs.ITcs.ETcs.LGeess.SPmath.IT

TL;DR

Resource-constrained edge devices may lack the memory and energy for digital neural-network inference, while wireless operators must also preserve data services. The paper jointly designs beamformers for multicast AI and unicast data flows under accuracy, energy, power, and hardware constraints; in representative 3GPP settings, the relative aggregate data rate reduction is below 5%.

  • Problem

    Resource-constrained edge devices may not have the memory and energy for digital inference, and integrated AI and data flows must share wireless resources that create interference.

    Source: §I · pp. 1–2

  • Method

    The paper models one multicast AI stream with multiple unicast data streams and jointly designs their beamformers to maximize sum rate subject to computing accuracy, client energy, BS power, and mixer constraints.

    Source: §I · p. 2

  • Results

    Less than 5% relative aggregate data rate reduction is reported in representative settings under 3GPP channels for integrated AI and data flows versus the data-flow-only reference.

    Source: §I · p. 2

  • Takeaways & Limitations

    Numerical results indicate telecom operators can multicast an AI model to multiple edge devices for on-device inference while largely preserving data communication performance.

    Source: §VI · p. 6

    The study considers one AI model at the BS; supporting multiple models, which would require multi-group multicast beamforming, is left for future work.

    Source: §I · p. 1

Abstract

from arXiv · show

Many resource-constrained edge devices require artificial intelligence (AI) services but are not powerful enough to store or execute neural networks (NNs). Analog radio frequency (RF) computing addresses this limitation by letting a base station (BS) multicast a waveform encoding NN weights to different clients. Each client then multiplies the received weight-encoded waveform with its locally generated input-encoded waveform through a passive mixer in its RF frontend, thereby realizing the matrix-vector multiplications (MVMs) needed for AI inference with an extremely low power and memory requirement. This can enable telecom operators to provide AI inference services to the edge devices connected to their networks. In practice, however, this multicast AI flow must share the spectrum, power, and spatial resources with other data flows. This paper considers a multi-user multiple-input multiple-output (MU-MIMO) downlink scenario and studies the resulting relative aggregate data rate reduction from integrating the multicast AI flow with the unicast data flows. We formulate a joint beamforming problem for the data and AI flows to maximize the aggregate data rate subject to computing accuracy, client energy, total power, and hardware-related constraints. We develop a weighted minimum mean-square error (WMMSE)-based algorithm to solve the problem. Numerical results based on 3rd Generation Partnership Project (3GPP) channels show that the relative aggregate data rate reduction from integrating an AI flow with data flows is below 5% in representative settings. This indicates that telecom operators can use analog RF computing to provide edge AI services while providing data communication services simultaneously.

I. INTRODUCTION

The introduction asks how telecom operators can add AI inference for resource-constrained devices while continuing data communication over shared wireless resources. It presents a joint-beamforming approach and reports that representative settings incur less than 5% relative aggregate data rate reduction.

I. INTRODUCTION

  • The motivation is to serve resource-constrained edge devices that may lack the memory and energy for digital on-device inference; offloading can add communication overhead and expose private information.The paper frames wireless-network inference as a way for telecom operators to offer AI services directly to these devices.Source: §I · p. 1
  • Analog RF computing uses a BS-multicast weight-encoded waveform and client-generated input waveforms, whose mixer multiplication realizes MVMs for inference.The paper describes this approach as requiring very low energy and memory consumption.Source: §I · p. 1
  • Sharing wireless resources creates interference in both directions: the AI stream interferes with data users, while data streams perturb AI clients’ MVM results.The paper therefore jointly designs the AI and data beamformers.Source: §I · p. 2
  • The paper models one multicast AI stream alongside multiple unicast data streams and compares their aggregate rate with a data-only reference under matched system conditions.The comparison holds data users, channels, bandwidth, and BS power the same.Source: §I · p. 2
  • A joint sum-rate maximization formulation accounts for MVM accuracy, client energy, BS transmit power, and mixer hardware constraints, and is addressed with WMMSE beamformer design.The WMMSE algorithm jointly designs the AI and data beamformers.Source: §I · p. 2
  • Less than 5% relative aggregate data rate reduction is reported in representative settings for integrated AI and data flows versus the data-flow-only reference.The result is based on 3GPP channels; the reduction depends on computing accuracy, spatial degrees of freedom, AI-client count, and client energy limit.Source: §I · p. 2

II. ANALOG RF COMPUTING FOR AI FLOW

Analog RF computing implements each neural-network layer’s MVM by mixing weight-encoded and input-encoded waveforms. Frequency assignment aligns products for each weight-matrix row at a distinct output tone, which is filtered and read out.

A. Mixer-Based MVM

  • The mixer multiplies time-domain weight-encoded and input-encoded waveforms; assigned tones produce sum- and difference-frequency components.Filtering removes unwanted products before readout.Source: A. Mixer-Based MVM · p. 2
  • Products across the columns of each weight-matrix row align at one output tone, enabling the mixer output to encode the MVM result.The input-tone frequency minus the weight-tone frequency places each product at −m∆f, independent of column index n.Source: A. Mixer-Based MVM · p. 2
  • A bandpass filter rejects unwanted components, and a DFT readout extracts the selected tones as the MVM output.If a layer has more output rows than the hardware supports simultaneously, its rows are processed in blocks.Source: A. Mixer-Based MVM · p. 2
  • The BS multicasts the NN weight matrices layer by layer, with the physical-layer design fixed within a layer and reconfigured between layers.The formulation therefore designs one layer at a time and omits the layer index.Source: A. Mixer-Based MVM · p. 3
  • For each AI client, the MVM amplitude contains a BS-controlled computing gain, while client scaling β_k makes the mixer output proportional to g_kβ_k.The AI beamformer and client scaling determine computing accuracy and client energy consumption.Source: A. Mixer-Based MVM · p. 3

B. Computing Accuracy Model

The computing-accuracy model treats receiver noise and data-flow interference as independent zero-mean errors that add to normalized MSE. It identifies how computing gain and client scaling affect these error terms.

B. Computing Accuracy Model

  • Receiver noise and data-flow interference contribute additively to the analog MVM’s normalized mean-square error as independent zero-mean errors.The passage identifies the computing-accuracy requirement using a root-NMSE target ε_k.Source: B. Computing Accuracy Model · p. 3
  • The model specifies data-flow interference at each AI client using the data-user beamformers f_D,j collected in F_D.Source: B. Computing Accuracy Model · p. 3
  • Increasing |g_k| suppresses both NMSE error terms, whereas increasing |β_k| reduces only the receiver-noise error.Source: B. Computing Accuracy Model · p. 3

C. Client Energy Consumption

Client energy is modeled per MAC as waveform-generation, sampling, and readout costs, with a target relative to digital inference. The WMMSE procedure searches feasible beamformer candidates and retains the one with the largest data rate.

C. Client Energy Consumption

  • Client energy per MAC combines input-waveform generation with fixed sampling and readout costs.The total energy for one MVM is the per-MAC energy multiplied by the number of MAC operations.Source: C. Client Energy Consumption · p. 3
  • The WMMSE algorithm searches target settings, updates data beamformers, and returns the feasible candidate with the largest data rate.Its iterations continue until the relative sum-rate change is below a tolerance.Source: C. Client Energy Consumption · p. 3
  • The client energy constraint sets analog energy relative to digital inference, while a smaller τE shifts more computing-gain demand to the BS.The client scaling factor is bounded by both the energy-derived limit and the RF input limit.Source: C. Client Energy Consumption · p. 3

III. INTEGRATED DATA AND AI FLOWS: SYSTEM MODEL AND PROBLEM FORMULATION

The system model has a BS serving AI clients through one multicast AI stream and data users through unicast streams. It characterizes data-user SINR and aggregate data rate while accounting for interference from the AI stream.

A. System Model

  • A fully digital BS with Nt antennas and Nt RF chains serves single-antenna AI clients through one multicast stream and single-antenna data users through unicast streams.The BS serves both groups in the same time-frequency resource.Source: A. System Model · p. 3
  • The model represents receiver thermal noise as complex Gaussian noise with variance determined by bandwidth, temperature, and noise figure.The noise variance is σ²j = kBT0B/ρnf.Source: A. System Model · p. 3
  • The data-user SINR expression accounts for interference from other data streams and the AI stream, as well as receiver noise.The passage identifies the AI-stream interference term in the SINR.Source: A. System Model · pp. 3–4
  • The model uses receiver thermal noise for each data user, with variance σ²j = kBT0B/ρnf.This noise term enters the data-user SINR.Source: A. System Model · p. 3
  • The AI stream interferes with data users, while data streams perturb AI-client MVM results.These interference directions arise because the streams share wireless resources.Source: A. System Model · p. 4
  • The communication objective is the aggregate data rate across the data users.The rate is based on the data-user SINRs.Source: A. System Model · p. 4

B. Problem Formulation

The formulation maximizes aggregate data rate while enforcing computing, power, and mixer constraints. It measures the AI flow’s rate cost against a matched data-only reference and includes a non-convex computing constraint.

B. Problem Formulation

  • For fixed BS beamformers, setting |βk| to its upper bound is optimal and removes the client-side variable from the computing constraint.Increasing |βk| monotonically reduces the noise term in the MVM expression.Source: B. Problem Formulation · p. 4
  • Received computing power is the BS-controlled received power of the weight-encoded waveform, constrained by accuracy, data-flow interference, and the mixer LO limit.The accuracy-scaled power must exceed aggregate data-flow interference by a threshold, while the LO port remains in its small-signal region.Source: B. Problem Formulation · p. 4
  • The objective is to maximize aggregate data rate subject to BS power, beamformer-only computing, and passive-mixer hardware constraints.The coupled SINRs and difference of quadratic terms make the problem non-convex.Source: B. Problem Formulation · p. 4
  • The relative rate reduction compares the integrated-flow rate with an optimized data-only reference using the same users, channels, bandwidth, and BS power.Only the integrated rate includes the AI flow and its constraints.Source: B. Problem Formulation · p. 4

IV. WMMSE-BASED JOINT BEAMFORMING

The joint-beamforming procedure searches target received computing powers, derives corresponding AI beamformers and interference budgets, then updates data beamformers using WMMSE. It retains the feasible candidate with the largest data rate.

IV. WMMSE-BASED JOINT BEAMFORMING

  • The method handles coupling between AI and data beamformers by searching target received computing powers and applying WMMSE updates to data beamformers.It retains the feasible candidate with the largest data rate.Source: §IV · p. 4

A. Search of Received Computing Power

  • The finite search varies each client’s target received computing power from its minimum toward the mixer LO threshold.At the minimum, the data-flow interference budget is zero; increasing the target permits more interference at the AI client but consumes BS power and can increase AI-stream interference at data users. Feasibility requires qk,min ≤ PLO,th.Source: A. Search of Received Computing Power · p. 4

B. Update of AI Beamformer

  • For each target setting, the AI beamformer is obtained from an equality-constrained design that penalizes its energy in the data-user channel subspace.The method sets the common phase reference to zero because the constraints depend on received magnitude and client scaling compensates phase.Source: B. Update of AI Beamformer · p. 4

C. Update of Data Beamformers

The data-beamformer update converts aggregate-rate maximization into a weighted MSE optimization, then solves for beamformers under power and AI-client interference constraints.

D. Computational Complexity

The data-beamformer update's cost is dominated by matrix factorization for each received-power candidate; the factorization is reused across data beamformers, and candidates can run in parallel.

V. PERFORMANCE EVALUATION

The evaluation uses a 3GPP indoor-factory channel and a representative LeNet layer. At the calibrated accuracy target, integrating the AI flow reduces aggregate data rate by 4.67%, while tighter targets incur larger rate costs.

V. PERFORMANCE EVALUATION

  • The evaluation uses 3GPP TR 38.901 indoor-factory channels, with Nt = 256, seven data users, four AI clients, and a representative second LeNet convolutional layer.Source: §V · p. 5
  • The default ǫ = 0.18 is the largest calibrated error tolerance maintaining at least 96% MNIST accuracy for the LeNet task.Source: §V · p. 5

A. Impact of Computing Accuracy Target

B. Impact of Number of AI Clients

The rate cost of integrating an AI flow depends on client count, spatial degrees of freedom, and client energy target. The paper concludes that representative settings can preserve data communication performance while supporting multicast AI inference.

B. Impact of Number of AI Clients

C. Impact of Spatial Degrees of Freedom

D. Impact of Client Energy Target

VI. CONCLUSION

  • The authors conclude that telecom operators can multicast an AI model for edge inference while largely preserving data communication performance; multiple AI flows, scheduling, mobility, and hardware validation remain future work.Source: §VI · p. 6

Related papers

Loading 2610.20841v1…