PaperlayerView exact revision on arXiv

Source-linked AI summary

Green Edge AI: A Contemporary Survey

Yuyi Mao, Xianghao Yu, Kaibin Huang, Ying-Jun Angela Zhang, Jun Zhang

arXiv:2312.00333v2cs.AIcs.ITcs.NImath.IT

TL;DR

Edge AI’s device-side learning and inference face energy and resource constraints. This survey analyzes energy components, derives design principles, and reviews mitigation strategies across data acquisition, training, and inference.

Abstract

from arXiv · show

Artificial intelligence (AI) technologies have emerged as pivotal enablers across a multitude of industries largely due to their significant resurgence over the past decade. The transformative power of AI is primarily derived from the utilization of deep neural networks (DNNs), which require extensive data for training and substantial computational resources for processing. Consequently, DNN models are typically trained and deployed on resource-rich cloud servers. However, due to potential latency issues associated with cloud communications, deep learning (DL) workflows are increasingly being transitioned to wireless edge networks in proximity to end-user devices (EUDs). This shift is designed to support latency-sensitive applications and has given rise to a new paradigm of edge AI, which will play a critical role in upcoming sixth-generation (6G) networks to support ubiquitous AI applications. Despite its considerable potential, edge AI faces substantial challenges, mostly due to the dichotomy between the resource limitations of wireless edge networks and the resource-intensive nature of DL. Specifically, the acquisition of large-scale data, as well as the training and inference processes of DNNs, can rapidly deplete the battery energy of EUDs. This necessitates an energy-conscious approach to edge AI to ensure both optimal and sustainable performance. In this paper, we present a contemporary survey on green edge AI. We commence by analyzing the principal energy consumption components of edge AI systems to identify the fundamental design principles of green edge AI. Guided by these principles, we then explore energy-efficient design methodologies for the three critical tasks in edge AI systems, including training data acquisition, edge training, and edge inference. Finally, we underscore potential future research directions to further enhance the energy efficiency of edge AI.

I. INTRODUCTION

Edge AI moves AI workloads closer to users to address cloud latency, but its energy demands strain battery-powered devices. This survey reviews energy-efficient approaches across edge-AI tasks, addressing a gap in holistic coverage.

A. Background; B. The Energy Challenge in Edge AI

B. The Energy Challenge in Edge AI

  • DNN growth drives substantial computation and energy demands, while edge-AI energy also comes from sensing and frequent wireless communication.The survey identifies limited battery capacity, resource constraints, and communication energy as central concerns for sustainable edge-AI services.Source: B. The Energy Challenge in Edge AI · p. 2

C. Related Survey Papers and Our Contributions

  • Prior surveys cover selected edge-AI topics, but lack a holistic review of energy-efficient methods for data acquisition, edge learning, and edge inference.This survey organizes those methods around energy-consumption components and design principles, and identifies future research opportunities.Source: C. Related Survey Papers and Our… · p. 3

D. Paper Organization

The paper proceeds from edge-AI task definitions and energy-design principles to reviews of energy-efficient methods for data acquisition, training, and inference. These tasks differ in where data and computation reside and in their dominant on-device energy costs.

D. Paper Organization

  • Sections II–V introduce edge-AI tasks and design principles, then review energy-efficient approaches for data acquisition, distributed training, and inference; Sections VI–VII cover research directions and conclusions.Figure 1 summarizes this organization.Source: D. Paper Organization · p. 3

II. KEY TASKS OF EDGE AI AND GENERAL DESIGN PRINCIPLES OF GREEN EDGE AI

  • The survey decomposes on-device energy into sensing, communication, and computation, then derives general design principles from the task-specific energy breakdown.Data acquisition mainly uses sensing and communication energy; distributed training and inference are dominated by communication and computation.Source: §II · p. 3, A. Key Tasks of Edge AI · p. 5

A. Key Tasks of Edge AI

  • Edge AI comprises centralized training with EUD data sent to MEC servers, distributed training that exchanges model updates, and inference using EUD and MEC resources.Distributed training avoids gathering local data at the MEC server, while inference may partition a DNN between an EUD and an MEC server.Source: A. Key Tasks of Edge AI · pp. 4–5

B. Breakdown of Energy Consumption

On-device energy varies by operation and platform: sensing depends on the sensor and sampling configuration, communication on the technology and transmission settings, and computation on the platform and workload. These differences motivate task-specific energy-efficient designs.

B. Breakdown of Energy Consumption

  • 96 mW for active imaging versus 165 μW in standby shows sensing power depends strongly on operating mode; imaging draws 96 mW versus 0.72 mW for heart-rate sensing.Sensing energy also varies with sensing mechanism, sampling rate, and sampling quality.Source: B. Breakdown of Energy Consumption · p. 5
  • Up to 15 dBm (∼32 mW) in 2.4 GHz WiFi and 23 dBm (∼199 mW) in 5G sub-6 GHz illustrate communication-power variation by technology.The passage also reports a 5G millimeter-wave maximum of 26 dBm (∼398 mW); communication power depends on factors including carrier frequency, modulation, data rate, and antenna deployment.Source: B. Breakdown of Energy Consumption · p. 5
  • Training ResNet-110 and VGG-16 on NVIDIA Jetson TX2 requires approximately 8 × 10^5 and 3.8 × 10^5 Joules, respectively.These correspond to approximately 17.5 and 8.3 times an iPhone 14 battery’s capacity; specialized TPUs are more energy-efficient than general-purpose GPUs.Source: B. Breakdown of Energy Consumption · p. 6
  • Because each edge-AI task has a distinct interplay of sensing, communication, and computation costs, energy-minimization approaches need to be customized to the task.The survey uses these components as the basis for its overarching energy-efficient design principles.Source: B. Breakdown of Energy Consumption · p. 6

C. General Design Principles

Green edge AI should balance intelligence and energy across sensing, communication, and computation while adapting operations to changing system conditions. For centralized edge learning, data-acquisition approaches target lower sensing and transmission costs or effective learning from smaller datasets.

C. General Design Principles

  • Energy-efficient design should jointly balance intelligence and energy across sensing, communication, and computation, supported by cooperation between EUDs and MEC servers.The survey defines energy efficiency as intelligence obtained per joule and calls for re-architecting networks to coordinate EUD and MEC resources.Source: C. General Design Principles · p. 6
  • Adapt sensing, transmission, and processing to changing scenes, channels, workloads, task states, and EUD populations to improve energy efficiency.Examples include adjusting vehicular LiDAR sampling to scene informativeness and adapting uplink schemes to instantaneous channel conditions.Source: C. General Design Principles · p. 6
  • When further intelligence gains require super-linearly more energy, application-specific systems should compute only what matters and collect just enough data for target accuracy.Task-oriented trade-off optimization can use prior information obtained through offline profiling or online learning.Source: C. General Design Principles · pp. 6–7

III. ENERGY-EFFICIENT DATA ACQUISITION FOR CENTRALIZED EDGE LEARNING

  • Centralized learning moves model training to MEC servers, but EUD sensing and data transmission remain major device-energy costs addressed through sampling, transmission, and small-dataset methods.The section asks how to sample and transmit efficiently, and how to improve learning when the MEC server receives little training data.Source: §III · p. 7

A. Adaptive Sampling

Adaptive sampling exploits changing data to reduce sensing and transmission costs, while learning-aware transmission prioritizes or compresses useful samples. Related approaches manage sample importance and dataset size to balance energy use and learning performance.

A. Adaptive Sampling

  • For WSNs, adaptive sensing cut sensing energy by up to 60%, while similarity search saved 93% transmission energy with <7% data-accuracy loss; MEC methods saved 12.86∼35.83% and up to 41%.The MEC methods adjust sampling frequency from recent data or select sensors using spatial and temporal correlations, with intervals adapted to temporal prediction error.Source: A. Adaptive Sampling · pp. 7–8
  • For task-oriented sampling, balances recognition error and sensing energy, while reduced trajectory data by ∼70% without losing critical queue-location points.The latter adjusts sampling frequency using vehicle-flow estimates and detected stop/go trajectory segments.Source: A. Adaptive Sampling · p. 8

B. Learning-aware Data Sample Transmission

  • On MNIST, evaluating only 2.5% of samples maintained similar learning accuracy; importance metrics include training loss, prediction uncertainty, and sample-to-centroid distance.Sample importance can change as the learned model is updated.Source: B. Learning-aware Data Sample… · pp. 8–9
  • Transmission can prioritize or compress samples, protect high-quality data with more resources, use importance-adaptive retransmission, and schedule devices toward target learning performance.Analog modulation reduces EUD baseband processing, while retransmission continues until effective SNR exceeds a threshold adapted to sample importance.Source: B. Learning-aware Data Sample… · pp. 8–9
  • Power-allocation studies use data-volume–training-loss power laws to set sample counts, with extensions for wireless resources, device scheduling, continuous arrivals, and MEC limits.The relationship between data volume and training loss is difficult to determine exactly, so data-driven models provide approximate characterizations.Source: B. Learning-aware Data Sample… · p. 9

C. Centralized Edge Learning with Small Datasets

For centralized edge learning with limited data, the survey reviews data augmentation and transfer learning as ways to support learning performance while reducing data-acquisition demands. It also identifies unresolved needs in energy modeling and in understanding combined techniques, and places cooperative training within the broader edge-AI setting.

1) Data Augmentation:

  • Data augmentation expands dataset size, quality, and diversity; cited applications suggest it can support target learning performance with less or lower-quality transmitted data, saving sensing and communication energy.The cited application studies did not directly address EUD energy consumption.Source: 1) Data Augmentation · p. 9

2) Transfer Learning:

  • Transfer learning reuses knowledge for related tasks; one study reported less training data and time with better accuracy, while another eliminated up to 70% of source images sent to target cameras.The component-recognition study transferred a pretrained VGG-16 model; the camera study used elite-instance-based matching to select images.Source: 2) Transfer Learning · p. 9

D. Takeaways

  • Data-acquisition methods need tractable links between design choices and learning performance, accurate sensing-energy models, stronger theoretical support, and investigation of minimum data needs with DA and TL.The survey notes that existing work is mainly heuristic and identifies offline data-fitting and online learning as ways to obtain performance models.Source: D. Takeaways · p. 10
  • The composite effects of the three reviewed technique classes on energy-efficient centralized data acquisition warrant further investigation because the classes were examined separately.Source: D. Takeaways · p. 10

IV. ENERGY-EFFICIENT COOPERATIVE TRAINING AT THE EDGE NETWORK

  • Cooperative edge learning distributes training across EUDs; the survey focuses on federated learning and says its energy-efficient approaches also apply to DEEL and H-FEEL.Cooperative training can leverage distributed data and computational resources where centralized transfer demands substantial bandwidth and may conflict with data-privacy regulations.Source: §IV · p. 10

A. Background of Federated Learning

FEEL moves federated learning to resource-constrained mobile edge networks, where computation and communication energy must be managed without substantially degrading learning. The survey reviews model and gradient compression, including quantization and sparsification, as energy-saving approaches.

A. Background of Federated Learning

  • FEEL clients download a global model, train locally, and upload updates for parameter-server aggregation, keeping training data local.This better protects privacy than centralized edge learning.Source: A. Background of Federated Learning · p. 11

1) Quantization:

  • [126] reports up to 70% energy savings over full-precision training without convergence-rate deterioration, selecting precision from local-training and transmission energy models.Quantization reduces parameter precision, but the irreversible process requires balancing energy efficiency with learning performance.Source: 1) Quantization · p. 11
  • QSGD reduces communication energy by 5.7× versus standard SGD, while sign-only gradient exchange cuts energy 18∼40% and improves accuracy versus SignSGD and FedAvg.AdaQuantFL gradually increases quantization levels, converging with less communication energy than fixed-level quantization at minor accuracy loss.Source: 1) Quantization · p. 11

2) Sparsification:

  • Gradient thresholding can reduce communication energy hundreds of times without compromising accuracy; combined quantization and sparsification saves 54.7∼79.5% on CIFAR-10.Fairness-aware sparsification improves accuracy by up to 40% over a given training period under statistical data heterogeneity.Source: 2) Sparsification · p. 11
  • PruneFL adapts model size during training; versus FedAvg, it saves 35.2% of FLOPs and 71.2% of training time while reaching 80% accuracy on FEMNIST [140].The model can grow or shrink to maximize empirical risk reduction divided by round duration.Source: 2) Sparsification · pp. 11–12

3) Knowledge Distillation:

Knowledge distillation can accommodate heterogeneous device resources, while FEEL resource management balances local computation and communication against energy use and learning performance. The survey reviews local-training adaptation and joint computation-communication control.

3) Knowledge Distillation:

  • FedMD supports heterogeneous EUD model architectures by exchanging aggregated logits, reducing communication overhead versus FedAvg; FedKD uploads smaller student gradients with marginally higher local computation energy.Mix2FLD performs logit-to-model conversion at the MEC server and reports faster convergence and higher accuracy, implying lower EUD energy consumption.Source: 3) Knowledge Distillation · p. 12

C. Resource Management

  • FEEL resource management seeks energy efficiency while preserving learning performance, adapting training parameters and allocating limited communication and computation resources.The reviewed perspectives include local training, joint computation-communication control, device selection, and data selection and offloading.Source: C. Resource Management · p. 12

1) Local Training Adaptation:

  • Local training trades higher per-iteration computation energy for potentially fewer iterations and lower total communication energy, so training effort must be adapted.Dynamic batch sizes and learning rates address heterogeneous EUDs; CPU-cycle frequency control reports over 4% energy savings and substantially reduced training time versus static frequency.Source: 1) Local Training Adaptation · pp. 12–13

2) Joint Computation-Communication Control:

  • Jointly optimizing local training and model uploading reduces energy versus equal-bandwidth and sub-optimal-training baselines in TDMA-based FEEL.The reviewed controls include CPU frequency, local training accuracy, transmit power, and bandwidth allocation; related investigations cover FDMA and NOMA systems.Source: 2) Joint Computation-Communication… · p. 13

3) Device Selection:

FEEL energy management includes choosing participating devices, selecting useful local data, and offloading data when privacy and resource conditions permit. The survey also introduces DEEL and H-FEEL as other cooperative edge-learning frameworks.

3) Device Selection:

  • Device selection must account for both learning performance and device energy, since wireless bandwidth limits how many EUDs can participate.Long-term selection can outperform strategies that myopically minimize each round's energy; AutoFL reports 5.2× EE improvement versus random device selection.Source: 3) Device Selection · p. 13

4) Data Selection:

  • Data selection can save local-computation and model-exchange energy by removing irrelevant samples or training on informative ones, but assessing sample utility is challenging when data stay local.A sample-importance method reports over 70% energy reduction on MNIST and CIFAR-10.Source: 4) Data Selection · pp. 13–14

5) Data Offloading:

  • When EUDs and MEC servers share an authority, data offloading can reduce on-device training overhead; raw-data offloading may be unsuitable under stringent privacy requirements.The survey also describes nearby-device offloading and privacy-preserving coded FL as an option for such cases.Source: 5) Data Offloading · p. 14

D. EE-oriented Designs for Other Cooperative Edge Learning Frameworks

  • The survey introduces DEEL and H-FEEL as alternative cooperative edge-learning frameworks and reviews their energy-efficiency-oriented designs.Their specific designs are covered in the following framework discussion.Source: D. EE-oriented Designs for Other… · p. 14

1) DEEL:

The survey reviews energy-efficient designs for DEEL and H-FEEL, then turns to edge inference paradigms and their energy approaches. It emphasizes that fully on-device inference faces computation, battery, and thermal constraints.

1) DEEL:

  • DEEL aggregates models at individual EUDs through one-hop neighbors, making it more resilient to node failures than FEEL; designs include quantized updates and joint pruning and topology construction.In a beat-signal dataset experiment, accuracy only slightly degrades with 1-bit quantization.Source: 1) DEEL · p. 14

2) H-FEEL:

  • H-FEEL uses clustered EUDs associated with MEC servers and two-level edge/cloud aggregation; device-edge association affects learning performance and cost.An association method minimizing divergence from a hypothetical central model can reduce computation and communication energy with fewer training rounds.Source: 2) H-FEEL · pp. 14–15

E. Takeaways

  • Green FEEL combines compression to reduce communication and local-training computation with resource management that adapts training, device, and data participation.The survey identifies device, model, and data heterogeneity, device failures, and large-scale EE analysis as practical research concerns.Source: E. Takeaways · p. 15

V. ENERGY-EFFICIENT EDGE INFERENCE: PARADIGMS AND ALGORITHMS

  • The survey contrasts on-device, MEC server-based, and device-edge cooperative inference; the first two respectively impose high on-device computation and communication overhead.The paradigms and corresponding energy-efficient design approaches are summarized in Figs. 6 and 7.Source: §V · p. 15, Figure 6 · p. 15, Figure 7 · p. 16

A. On-device Inference

  • On-device inference suits EUDs with reasonable compute resources and privacy-sensitive applications, but DNN processing is constrained by computation, battery energy, and thermal limits.Prior efforts mainly targeted real-time performance without scrutinizing energy footprint.Source: A. On-device Inference · p. 16

1) Model Compression:

On-device inference efficiency is pursued through model compression, energy-aware architecture search, and input-adaptive computation. These approaches reduce computation or memory costs while balancing energy use against accuracy.

1) Model Compression:

  • Quantization, sparsification, and distillation reduce inference costs through lower-precision arithmetic, fewer operations or memory accesses, and smaller student models.FitNets reduces parameters 10× relative to a larger teacher while improving inference accuracy; progressive distillation permits runtime selection among small students.Source: 1) Model Compression · pp. 16–17
  • HAQ saves 50% of inference energy with < 0.5% top-1 accuracy drop on ImageNet versus 8-bit MobileNet-V1.Its hardware simulator informs a deep-reinforcement-learning agent selecting mixed precision.Source: 1) Model Compression · p. 16
  • Energy-aware pruning achieves 3.7× energy reduction for AlexNet and 1.6× for GoogleNet, each with < 1% top-5 accuracy loss on ImageNet.Another method removes ~50% of FLOPs with minor accuracy reduction by pruning during training.Source: 1) Model Compression · pp. 16–17

2) Neural Architecture Search:

  • Mixed-precision NAS reduces inference energy by 20.9% versus HAQ and 76.2% versus MobileNetV2 on CIFAR-100, with similar inference accuracy.Energy-aware NAS can measure candidate energy using profiling tools or estimate it with prediction models.Source: 2) Neural Architecture Search · p. 17

3) Input-Adaptive Inference:

  • Input-adaptive inference adjusts computation to input difficulty through model selection, early exits, or layer and channel skipping.Model selection routes inputs among models; early termination uses intermediate predictions; skipping uses learned gates or policies, whose overhead should remain negligible.Source: 3) Input-Adaptive Inference · p. 17, Figure 8 · p. 18

B. MEC Server-based Inference

MEC server-based inference shifts DNN execution to the server, where transmitting inference data can dominate energy use. The reviewed approaches jointly consider resource allocation and data configuration while accounting for accuracy and latency.

B. MEC Server-based Inference

1) Communication-Computational Resource Management:

  • Offloading decisions must account for inference accuracy as well as latency and energy, including data quality and DNN computation demand.One approach estimates these factors from validation data; another jointly optimizes model placement and resource allocation under latency requirements.Source: 1) Communication-Computational… · p. 18

2) Data Preprocessing:

  • Preprocessing can configure inference data to reduce communication overhead, but its settings must account for effects on inference accuracy.Sampling rate, lossy compression, and video frame quality are examples of configurable data properties; some approaches model accuracy offline, while others adapt configurations through learning.Source: 2) Data Preprocessing · pp. 18–19

C. Device-edge Co-inference

Device-edge co-inference partitions DNN execution between an end-user device and an MEC server, trading communication against computation energy for a target accuracy. Energy efficiency also depends on feature compression and cooperation among devices.

C. Device-edge Co-inference

  • Co-inference splits a backbone DNN between the EUD and MEC server, with the device sending an extracted feature map for server-side completion.The partition point controls the communication-computation energy trade-off at a target inference accuracy.Source: C. Device-edge Co-inference · p. 19

2) Feature Compression:

  • Partitioning trades communication and computation energy for target accuracy, and some partition points amplify intermediate data, motivating feature compression.Lossless compression achieves a 1.5–3× compression ratio; lossy compression removes more redundancy with some accuracy loss.Source: C. Device-edge Co-inference · p. 19, 2) Feature Compression · p. 20

1) Adaptive Model Partitioning:

  • Neurosurgeon selects a runtime model partition point and reduces EUD energy consumption by 59.5% versus MEC server-based inference.Other adaptive schemes combine partitioning with early termination and select configurations according to network status or system dynamics.Source: 1) Adaptive Model Partitioning · pp. 19–20

3) Multi-device Cooperation:

  • CoEdge distributes image patches across cooperating EUDs and reduces energy consumption by 25–67% across four DNN models versus equal workload partitioning.Distributed multi-view inference also reduces communication cost 20× versus offloading raw images in a multi-view classification task.Source: 3) Multi-device Cooperation · p. 20
  • A two-EUD, two-MEC-server prototype saves more than 40% of mobile energy consumption and inference latency versus its baseline scheme.A separate multi-device scheme saves up to 56% of inference latency and 72% of energy consumption.Source: 3) Multi-device Cooperation · p. 20

D. Takeaways

The paper argues that edge inference paradigms should coexist for heterogeneous applications, with energy-efficient designs balancing computation, communication, and accuracy. It identifies sensing, hardware/software co-design, neuromorphic computing, green energy, and generative AI as future research directions.

D. Takeaways

  • On-device, MEC server-based, and device-edge co-inference each have distinct energy levers, and future systems are expected to combine them across heterogeneous applications.The paper identifies unified design methods, energy scaling laws, and sensing overhead as open considerations.Source: D. Takeaways · p. 20

VI. FUTURE RESEARCH DIRECTIONS

  • The future-directions agenda includes low-power sensing, hardware/software co-design, neuromorphic computing, green energy-powered edge AI, and energy-efficient mobile-edge GenAI.The neuromorphic topics include spiking neural networks and compute-in-memory techniques.Source: §VI · pp. 20–21

A. Low-Power Sensing via ISAC

  • ISAC can use radio propagation for sensing and relieve EUDs of power-intensive data acquisition, but full-duplex MEC introduces self-interference challenges.EUD sensing over uplink or sidelink signals also requires waveform and beamforming design that balances sensing accuracy, communication rate, and intelligence performance.Source: A. Low-Power Sensing via ISAC · p. 21

B. Hardware/software Co-design

  • Hardware/software co-design matters because energy efficiency depends on hardware implementation, and hardware-aware NAS results may be limited to fixed configurations.Expanding hardware search spaces raises search-cost challenges; training also needs specialized co-design because its compute and memory patterns differ from inference.Source: B. Hardware/software Co-design · p. 21

C. Neuromorphic Computing

The survey reviews neuromorphic computing and green-energy-powered approaches to edge AI, including their efficiency opportunities and practical constraints. It also considers green generative edge AI and identifies research directions for sustainable edge intelligence.

C. Neuromorphic Computing

  • SNNs can be orders of magnitude more energy-efficient than DNNs, but their advantage varies by inference task and they pose training and hardware challenges.An automotive voice-control demonstration reported 1000× energy efficiency and 200 ms latency improvement over a GPU-based CNN while maintaining comparable inference accuracy.Source: C. Neuromorphic Computing · pp. 21–22
  • Compute-in-memory can perform MAC operations within memory subarrays, but analog errors and immature software remain limitations.Mythic’s announced flash-memory processor delivered 25 TOPS at 3 Watts, reported as one-tenth of desktop GPU power consumption.Source: C. Neuromorphic Computing · p. 22

D. Green Energy-powered Edge AI

  • Green-energy-powered inference faces interruptions from variable energy availability; task-based execution and energy-aware multi-exit selection are explored, while MEC co-inference remains uncharted,,.The multi-exit approach selects an exit at runtime using Q-learning according to available energy.Source: D. Green Energy-powered Edge AI · p. 22
  • Green-energy-powered FEEL remains underexplored, with general convergence behavior still unclear and joint learning- and energy-aware resource management strategies pending,.Unexpected device dropout can make convergence difficult for conventional FEEL algorithms such as FedAvg; one study addresses this for strongly convex loss by scaling the learning rate with energy-arrival probability.Source: D. Green Energy-powered Edge AI · p. 22

E. Green Generative Edge AI

  • Generative edge AI must fit compute-intensive foundation-model workloads to resource-limited MEC environments while reducing energy use.The survey discusses parameter-efficient fine-tuning, split learning, model compression, device cooperation, and energy-efficient SNNs as possible approaches,, [326],, [328].Source: E. Green Generative Edge AI · pp. 22–23

VII. CONCLUSIONS

  • The survey synthesizes energy-efficiency strategies and highlights adaptability and balancing intelligence with greenness as design principles, alongside future research directions for sustainable 6G intelligence.Highlighted directions include integrated sensing and communication, hardware/software codesign, neuromorphic computing, green-energy-powered edge AI, and green generative edge AI.Source: §VII · p. 23
Loading 2312.00333v2…