Disaggregated Deep Learning via In-Physics Computing at Radio Frequency

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.

Disaggregated Deep Learning via In-Physics Computing at Radio Frequency

The paper presents a unique approach to disaggregated deep learning (DL) using in-physics computing at radio frequencies, proposing a novel architecture termed {} (WIreless Smart Edge networks). This edge computing architecture is designed to enhance energy efficiency and facilitate simultaneous deep learning (DL) model access for multiple clients in wireless edge networks. It leverages the intrinsic analog multiplication capabilities of frequency mixers to realize matrix-vector multiplications (MVMs) at radio frequency, bypassing conventional digital computation constraints.

Architecture and Methodology

In {}, model weights are frequency-encoded and broadcast by a central radio using a shared wireless channel. Clients receive these broadcasts and employ a passive frequency mixer to locally perform MVMs required for deep learning inference. This method minimizes data movement, addressing one of the significant energy efficiency bottlenecks in modern DL, especially as models scale to billions of parameters.

Key components of the architecture include:

• Central Radio: Transmits model weights encoded on radio-frequency waveforms, employing orthogonal frequency-division multiplexing (OFDM) and I/Q modulation to maximize spectral efficiency.
• Edge Client: Uses a frequency mixer to multiply the received model weights with locally encoded inference requests. This multiplication occurs directly at RF, achieving an analog MVM which facilitates DL inference.

Significant emphasis is placed on mitigating distortive effects introduced by wireless transmission. Techniques like channel state information (CSI) estimation and weight precoding are employed to preserve computation integrity over the shared channel.

Experimental Results and Analysis

Experiments using a software-defined radio (SDR) platform demonstrate the efficacy of the proposed method. Notably, {} achieves an energy efficiency of {6.0}\thinspace{fJ/MAC} with a classification accuracy of {95.7\%} on the MNIST dataset, improving computational efficiency to {165.8}\thinspace{TOPS/W}. These results indicate a substantial enhancement over traditional digital ASICs, which operate at {1}\thinspace{pJ/MAC}. The efficiency further improves under optimized conditions, underscoring the potential scalability and practical applicability of the system.

Moreover, the architecture's ability to approach thermodynamic limits in energy efficiency, even surpassing Landauer's bound under ideal scenarios, is emphasized as a compelling advantage in computations involving extensive MVMs. This is achieved through a reduction in active hardware components at each client and utilizing RF mixers inherently designed for signal multiplication.

Implications and Future Prospects

The paper delineates significant implications for the future of edge computing and in-physics DL. By offloading substantial computational tasks to the analog domain, this architecture could pivotally transform wireless edge networks, potentially integrating into a wide array of real-world applications.

The scalability aspect is discussed, with prospects of employing larger bandwidths and more advanced hardware, like beamforming antenna arrays and enhanced RF electronics, to further broaden the deployment scope and enhance computation throughput.

In terms of broader impacts, the findings suggest considerable future developments in low-energy AI systems, particularly in resource-constrained environments or applications employing large-scale DL models like convolutional neural networks and transformers. The inherent privacy benefits, courtesy of disaggregated computations, could also be pivotal in environments where data integrity and security are paramount.

In conclusion, the paper lays a robust framework for integrating in-physics computing at RF into AI and DL, presenting a promising direction to alleviate current limitations related to energy efficiency and data movement in complex neural networks.