Disaggregated Deep Learning via In-Physics Computing at Radio Frequency

TL;DR

The paper introduces WISE, a disaggregated deep learning framework that performs in-physics, RF-based matrix–vector multiplications by broadcasting frequency-encoded model weights to edge clients. Leveraging OFDM, I/Q modulation, and passive RF mixers, WISE achieves energy-efficient inference with orders-of-magnitude improvements over digital ASICs, e.g., up to 165.8 TOPS/W at 95.7% MNIST accuracy, and approaches thermodynamic limits for large problem sizes. It develops multiple channel-calibration schemes (W-precoding and x-precoding) to compensate for wireless distortions, analyzes energy efficiency and computation throughput, and validates the approach on software-defined radios with real DL tasks (MNIST and AudioMNIST). The work suggests that wireless-disaggregated DL, powered by in-physics RF computation, can dramatically reduce energy per MAC and enable scalable DL deployment on edge devices, with potential extensions to wired links and larger bandwidth scenarios. Overall, WISE presents a practical pathway to near-thermodynamic-energy-efficient DL inference across wireless edge networks, leveraging existing RF front-ends and OFDM signal processing.

Abstract

Modern edge devices, such as cameras, drones, and Internet-of-Things nodes, rely on deep learning to enable a wide range of intelligent applications, including object recognition, environment perception, and autonomous navigation. However, deploying deep learning models directly on the often resource-constrained edge devices demands significant memory footprints and computational power for real-time inference using traditional digital computing architectures. In this paper, we present WISE, a novel computing architecture for wireless edge networks designed to overcome energy constraints in deep learning inference. WISE achieves this goal through two key innovations: disaggregated model access via wireless broadcasting and in-physics computation of general complex-valued matrix-vector multiplications directly at radio frequency. Using a software-defined radio platform with wirelessly broadcast model weights over the air, we demonstrate that WISE achieves 95.7% image classification accuracy with ultra-low operation power of 6.0 fJ/MAC per client, corresponding to a computation efficiency of 165.8 TOPS/W. This approach enables energy-efficient deep learning inference on wirelessly connected edge devices, achieving more than two orders of magnitude improvement in efficiency compared to traditional digital computing.

Figures (24)

• Figure 1: The WISE architecture enables disaggregated model access and energy-efficient deep learning (DL) to multiple clients in wireless edge networks.a, A central radio broadcasts frequency-encoded model weights, $\mathbf{W}$, onto a radio-frequency (RF) signal at the carrier frequency $F_{w}$, which is precoded to $\mathbf{V}$ to mitigate the distortion introduced during propagation over the wireless channel, $\mathbf{H}$. b, Each client equipped with a WISE-R encodes the inference request $\mathbf{x}$ at the carrier frequency $F_{x}$, and performs local DL inference for $\mathbf{y}$ at the carrier frequency $F_{y}$, where the matrix-vector multiplications (MVM), or essentially the fully connected (FC) layers, are realized using a passive frequency mixer. c, Illustration of the in-physic MVM computation during frequency down-conversion with frequency-encoded $\mathbf{W}$, $\mathbf{x}$, and $\mathbf{y}$.
• Figure 2: WISE's workflow with one central radio and multiple clients.a, Experimental setup for WISE using a software-defined radio (SDR) platform: b, A central radio simultaneously provides disaggregated deep learning (DL) model access to three edge clients, each equipped with a WISE-R. c, On each client, the computing mixer performs general matrix-vector multiplications (MVMs) in-physics using the wirelessly received model weights ($\mathbf{W}$) and local inference request ($\mathbf{x}$). d, The model weights $\mathbf{W}$ is modulated at $F_{w}={0.915}\textrm{GHz}$ over a wireless channel, and the inference request $\mathbf{x}$ is modulated at $F_{x}={1.2}\textrm{GHz}$; after down-conversion, the MVM result $\mathbf{y}$ is located at 0.285GHz. e, WISE achieves classification accuracies of 97.1%--97.4% across the three clients on the MNIST dataset using the LeNet-300-100 model, which is comparable to the accuracy of 98.1% achieved by traditional digital computing but with significantly improved energy efficiency.
• Figure 3: Benchmarking general complex-valued inner-product (IP) computation: computing accuracy and energy efficiency.a, Complex-valued IP computation of two length-$N$ vectors, $c = {\langle{\mathbf{a}},{\mathbf{b}}\rangle} = \sum_{n=1}^{N} a_{n} \cdot \overline{b_{n}}$, where $\mathbf{a}$ and $\mathbf{b}$ are frequency encoded onto $N$ (4,096) subcarriers across a bandwidth of $B$ (25MHz). b, Decoding of the IP result, $c$, after the in-physics IP computation, low-pass filtering, and sampling using an analog-to-digital converter (ADC). c, IP computing accuracy achieved by WISE as a function of the signal-to-noise ratio (SNR) for $N = 4,096$ and $N = 32,768$. d, Energy efficiency of WISE, $e_{\textrm{mvm}}$ (J/MAC), required to achieve $\textsf{RMSE} < 0.0625$ (equivalent to 5-bit computing accuracy sludds2022delocalizeddavis2022frequency) as a function of the IP size, $N$.
• Figure 4: WISE for energy-efficiency deep learning (DL) inferences.a/d, Deployment of WISE for DL tasks using complex-valued three-layer models: classification of handwritten digits on the MNIST dataset (a) and spoken digits on the AudioMNIST dataset (d). b/e, Experimental classification accuracy achieved by WISE on the MNIST (b) and AudioMNIST (e) datasets over different energy efficiency (J/MAC), and the corresponding energy consumption per inference. c/f, Confusion matrices for classification accuracy at 10dB and 20dB SNR on the MNIST (b) and AudioMNIST (e) datasets.
• Figure S1: The diagram of I/Q modulation and demodulation.a, The I/Q modulation converts the complex-valued baseband I/Q sequence $\mathbf{s}$ into a real-valued waveform $r(t)$ modulated at carrier frequency $F$. b, The I/Q demodulation converts the received real-valued waveform $r(t)$ back to the complex-valued baseband I/Q sequence $\Tilde{\mathbf{s}}$, which is proportional to the original $\mathbf{s}$.