WorksComputer Science
Energy-Efficient Timely Relevant IoT Data Retrieval with Wireless Memory Approximation
Abstract
Collecting relevant data from Internet of Things devices under freshness and energy constraints is a key challenge in 6G edge networks. A lightweight neural network on the device filters irrelevant observations, but storing it in dynamic random access memory consumes refresh energy during standby. This letter proposes a pull-based framework where an edge server remotely activates and approximates memory and jointly selects the activation timing, refresh period, model size, and transmission threshold. We derive the expected freshness and energy over erasure channels in closed form and show that the scheme consumes less energy than an always-on baseline under the same freshness constraint.
Introduction
Internet of Things (IoT) networks collect multimodal sensory data to support emerging applications such as wildlife monitoring [1] and the remote control of mobile robots [2]. Because such data has complex patterns, a lightweight neural network (NN) on the IoT device can process observations locally and transmit only the data relevant to a given task, which reduces transmission energy [3], [4]. The computation cost this adds is incurred only per inference and is offset by the transmission cost it saves. Keeping the model in dynamic random access memory (DRAM), by contrast, consumes energy continuously through periodic refresh, even during standby [5]. This standby refresh cost has been overlooked in protocol design for 6G applications.
DRAM approximation lengthens the refresh period to reduce this refresh energy at the cost of a controlled level of bit errors [6, 7, 8]. Existing schemes adapt the refresh rate to data criticality [6] or to a user-specified quality level in smart-camera and multiview inference systems [9], [10]. However, none of them adapts the approximation to the receiver’s time-varying goals or to the relevance of the stored model to the current task, so the device keeps refreshing a model that no task requires.
To address this, the framework of wireless memory activation and approximation [5] lets an edge server (ES) remotely activate DRAM only when a relevant model is needed and set its refresh period according to the required approximation level. The device then spends refresh energy only when a task requires the model, and only on that model. This framework, however, does not address the energy-efficient retrieval of timely relevant data under a deadline. In particular, it accounts for neither the semantic relevance filtering performed at the device nor the model size.
This letter studies energy-efficient retrieval of timely and relevant data from a single IoT device by a deadline, a problem central to decision-making in applications such as rescue robotics and wildlife monitoring. We adopt the framework of [5] and extend it as follows: ) we design a new energy-efficient communication framework and protocol that integrate memory, computation, and communication; ) we introduce an analytical model of the energy consumption of the IoT device and adopt a goal-oriented metric, the Grade of Effectiveness (GoE) [11], which captures the Age of Information (AoI) [12] of the freshest relevant observation at the deadline; ) we derive the expected GoE and energy consumption in closed form, jointly select the activation timing, refresh period, model size, and threshold for semantic filtering under a freshness constraint, and quantify the energy savings over an always-on baseline. This work is the first to tailor wireless memory activation and approximation to timely relevant data retrieval and to design an energy-efficient protocol that accounts for the interaction among memory, computation, and communication. Especially, compared with [5], this work newly investigates optimal selection of NN and threshold of semantic filtering for wireless memory approximation considering trade-off between energy consumption and information freshness with respect to the deadline based on our derived equations.
System Model
We consider a scenario in which an ES retrieves data from a single IoT device through a semantic query [1], which requires on-device inference. Figure 1 shows a configuration of the IoT device considered in this paper. The device has a vision sensor, a radio, non-volatile storage that holds a set of task-specific models of different sizes, and volatile DRAM that holds the active model for on-device inference with on-chip static random access memory (SRAM).
Figure 1. A configuration of the IoT device under wireless memory activation and approximation.
Figure 2(a) shows the pull-based communication setup. Each episode is triggered by a query request from the cloud and occurs at random, infrequent instants. The request specifies the content of interest (e.g., a desired bird), a deadline , and a freshness requirement . We focus on a single episode, which starts at . During the episode, the vision sensor samples the environment with period , which exceeds the time to transmit a single image.
Figure 2. (a) Communication setup; (b) timely relevant data retrieval from a single IoT device. Images , , and are relevant, and is the freshest at .
To reduce the refresh energy, we adopt the wireless memory activation and approximation framework of [5], illustrated in Figure 2(b). When a new episode begins, the ES decides when to transmit the memory activation signal based on the deadline , and sends it seconds after the episode begins. The signal identifies the relevant model by embedding its identity (ID). If the device detects the signal and the embedded ID matches one stored locally, it loads the corresponding model from storage into DRAM and begins processing its new observations, so DRAM is activated on demand and consumes no refresh energy before activation. The ES then sets the refresh period to , where a longer saves refresh energy at the cost of greater degradation of the model held in DRAM.
After memory activation, the ES transmits the threshold and a semantic query, which embeds the desired content as a feature vector , where is a query example and is a reference model at the ES. On receiving the query, the device evaluates the relevance of each new observation with the activated model and transmits the observation if its similarity measure is at least , as detailed in Sec. II-A2. We assume the downlink (DL) and uplink (UL) channels are erasure channels, in which packet loss occurs with probability and , respectively, without retransmission within an episode. We denote by the total overhead from the activation signal to the start of processing, which comprises the DL reception and the model loading.
Observation Model
Let and be the numbers of images that the IoT device captures before and after the memory activation, respectively, given by and . The total number of images in the episode is . The images captured after memory activation are denoted by . Here, is the -th image, where , , and are the number of channels, height, and width, respectively.
DRAM Retention Accuracy: Because DRAM cells store data as charge on leaking capacitors, DRAM must be refreshed periodically to retain the data. Let denote the nominal refresh period, which retains the data without error.
For , bit errors may occur in each refresh interval, as measured in [7]. Following [5], we model the bit error probability of a single refresh interval as , where is a hardware-specific constant, and a smaller gives faster memory degradation. The number of refresh operations between memory activation and the -th sample, denoted by , is
Then, the bit error probability after refresh operations is .
Similarity Measure: To infer the relevance of each observed image to the current task, the IoT device computes a similarity measure with the active model , as in [1], [3], [4]. We assume is a feature extractor that maps an image to a feature vector . Let be a similarity function. Then, the observed similarity of the -th image , denoted by , is . The observed similarity deviates from the true similarity obtained using . We assume that all device-side models have an embedding dimension of and their outputs are aligned with the query embedding space.
This deviation has two sources, the limited size of the device model and the DRAM degradation under memory approximation [5], and we model the observed similarity of the -th sample as
where , and the total variance combines a static term due to the limited model capacity and a time-varying term due to memory degradation. Our experiment with the ImageNet validation set [13] and MCUNet [14] shows that the observed error approximately follows a Gaussian distribution whose variance increases under pruning and bit errors. We assume that is independent across samples and independent of . In practice, bit errors persist in the stored weights, and the errors of successive samples are positively correlated. Missed detections then cluster, and the product forms of and in Sec. III give an optimistic GoE.
We model with empirical neural scaling laws [15], under which the reducible loss of a deep learning model decays as a power law in the model size and the data size. We neglect the data-dependent term, since the deployed model is trained on a large dataset, and model the capacity noise in terms of the number of parameters as , where is a scaling coefficient and is the scaling exponent.
Next, we model . After refresh operations, each stored bit is in error with probability . The bit errors perturb the quantized weights of the NN and hence the output feature vector, and we model the resulting variance as , where is an architecture-specific constant.
Grade of Effectiveness
To quantify the freshness of the relevant data at the deadline, we adopt the GoE [11], which measures the effectiveness of updates at the receiver. To define the GoE, we first extend the standard AoI metric [12]. Let and denote the reception and generation times of the -th received image, and let be the generation time of the latest image received by time whose true similarity is at least , i.e., . The AoI of relevant data at time is then , and, following the Query Age of Information (QAoI) [16], we define the GoE as . If the ES receives no relevant image by the deadline, we set , as in [17].
Energy Model
Computation
The NN inference energy is the sum of the DRAM access energy () and on-chip processing energy () [18]. Here, denotes the energy per -bit DRAM access.
At full precision, the read/write energy of a small local SRAM or register file equals the energy of a single multiply-accumulation (MUAC) operation, and an access to the main SRAM costs [18]. Then, , where , , and are the compute, weight-access, and activation-access energies, respectively. Here, is the inference complexity in terms of MUAC operations and is the total number of activations across the model, and , where is the full-precision bit width.
Communication
Let denote the set of images that pass the filter. With transmit and receive powers and , the communication energy is . Here, and are the UL and DL transmission times, and and are the corresponding data rates.
Memory Refresh
The energy of a single refresh follows from its timing overhead and the DRAM current (IDD) values [19]. Let be the number of refresh operations between memory activation and the deadline. Then, the total refresh energy is , where is the number of parameters of NN , while , and respectively represent the energy consumed by a single Auto Refresh (AR) command, that of a single row refresh, and the number of bits stored per DRAM row. Loading the selected model from non-volatile storage into DRAM consumes power over the loading time , so the loading energy is .
Analysis
The IoT device transmits the -th sample with probability
where and is the probability density function of the true similarity. Then, the expected total energy consumption is:
where .
The expected GoE follows from the Probability Mass Function (PMF) of the index of the freshest relevant sample that the ES receives. The -th sample is relevant and delivered to the ES with probability
Sample is the freshest relevant sample at the ES if it is relevant, successfully delivered, and none of the later samples is both relevant and successfully delivered, so this event has probability . The ES receives no relevant sample, and , with probability . Finally, the expected GoE is:
Using (3) and (5), we minimize the expected total energy consumption of the IoT device subject to the GoE constraint :
We solve (6) by grid search, in which we vary s with a 50 s step subject to , ms with a 32 ms step, with a 0.04 step, and .
| Parameter | Value | Parameter | Value |
| s | 10 s | ||
| 2000 | 0 | ||
| bps | 170/160 mW | ||
| (3, 128, 160) | 8/16 | ||
| 67.3 M | 0.325 M | ||
| 1.56 pJ | 1.85 pJ | ||
| 3.70 pJ | 236.8 pJ | ||
| 64 ms | 640 ms | ||
| 41.5 nJ | 157 nJ | ||
| 8 Kbit | 0.2 | ||
| 0.35 | 197.6 |
Table I. Default parameters for numerical evaluations.
Numerical Evaluation
We present results for GoE and energy consumption obtained from analysis and Monte Carlo simulation over trials, shown by lines and markers, respectively. As a baseline, we consider a semantic query with always-on memory, in which DRAM is always active with the refresh period and holds the model selected for the query before the episode begins, so the baseline incurs no overhead (). The baseline uses the threshold and model size that minimize its energy consumption under the GoE constraint.
Table I lists the default parameters, which follow [20] for the radio, [18] for the computation, with the energies scaled to , [19] for the DRAM refresh, and [14] for the NN. We take from [21] and set so that at . For simplicity of analysis, we assume that is uniform on and set . A non-uniform changes only the integrals in and .
Figure 3 shows the GoE and total energy consumption of our proposed scheme as a function of memory activation time , where we set . The analytical and simulation results agree closely for both GoE and energy consumption, and this agreement validates the analysis. The activation time trades GoE against energy consumption. An earlier activation (small ) gives the device more samples in which to detect and deliver a relevant image, so the GoE decreases, but the communication, computation, and refresh energy increase because all three are proportional to the active duration . A later activation (large ) saves energy but increases the GoE, because the ES more often collects no relevant image. A higher passes fewer images through the filter and lowers , so the GoE increases and the transmission energy decreases.
Figure 3. GoE and energy consumption as a function of .
Figure 4 shows GoE and total energy consumption as a function of the number of parameters , where we set ms. As increases, the GoE decreases and the total energy consumption increases. A larger model reduces and hence the probability of missed detection, but it occupies more DRAM rows and requires more DRAM accesses per inference, so the refresh and inference energy increase. A longer reduces the number of refresh operations, and the slope of the energy in decreases roughly as . The longer also increases , which limits the benefit of a larger model: falls below 0.01 at M, so for the degradation term dominates and the GoE flattens, whereas for the degradation term becomes zero and the GoE keeps decreasing.
Figure 4. GoE and energy consumption as a function of model size.
Figure 5 shows GoE against packet error probability, in which we change the value of or . The GoE increases with both error probabilities, but for different reasons. A DL error prevents the device from receiving the activation signal and the query, so the episode incurs the penalty , and by (5) the GoE grows linearly in . A UL error loses a single image, and an earlier relevant image can still reach the ES, so the GoE grows gradually with . The DL control signal therefore requires stronger protection than the UL data.
Figure 5. GoE as a function of UL/DL packet error probability.
Figure 6 shows the GoE and energy consumption as a function of . A larger shortens the interval between model loading and the deadline, so the device captures fewer images after activation. The GoE therefore increases, and the inference, transmission, and refresh energy decreases. In contrast, the loading energy increases with , and a larger increases the total energy.
Figure 6. GoE and energy consumption as a function of .
To evaluate the energy efficiency of our proposed scheme, we consider the following configurations : setting 1 (3200, 2000, 0, 0); setting 2 (320, 2000, 0, 0); setting 3 (3200, 5000, 0, 0); setting 4 (3200, 2000, 0, 0.1); setting 5 (3200, 2000, 0.1, 0). We obtain the optimal sets of parameters by solving Eq. (6). Then, we define , as the ratio of the total energy consumption of the proposed scheme to that of the baseline.
Figure 7 shows against for the different settings. Missing points indicate that no configuration satisfies the constraint. From Figure 7, our proposed scheme consumes less energy than the baseline when for settings 1 to 4. The ES achieves this saving by tuning the memory activation timing and refresh period under the GoE constraint. Setting 2 yields a higher than setting 1 at every because its smaller accelerates memory degradation and forces the ES to choose a shorter refresh period . In setting 3, the larger penalty for failing to collect relevant data forces the ES to activate the memory earlier, so also exceeds that of setting 1. In setting 4, UL losses reduce by the factor , so the ES must activate the memory earlier or lower to deliver at least one relevant image by the deadline, and this conservative choice keeps above that of setting 1 at every . The impact of channel error is larger for the DL signaling (setting 5). In setting 5, Eq. (5) gives , so no configuration satisfies . For larger , the budget left for missed detections shrinks to , and the ES must apply a smaller , a larger , a smaller , and an earlier , which increases energy consumption.
Figure 7. Energy consumption ratio against the GoE constraint .
Conclusion
This letter studied energy-efficient, timely, and relevant data retrieval from an IoT device under wireless memory activation and approximation. We derived closed-form expressions for the expected energy consumption and the GoE. The numerical results characterized the energy-freshness trade-off and showed that the proposed scheme consumes less energy than the always-on baseline for relaxed freshness constraints. Future work includes an experimental feasibility study and a robust design that accounts for mismatched statistics and temporally correlated observations.
References
- [1]A. E. Kalør, P. Popovski, and K. Huang, “Data sourcing random access using semantic queries for massive IoT scenarios,” IEEE Trans. Commun., vol. 73, no. 12, pp. 14 381–14 396, 2025.
- [2]R. Bezerra, S. Tadokoro, and K. Ohno, “AI-IoT-robotics integration: Survey of frameworks, emerging trends, and the path toward connected robotics,” IEEE Internet Things J., 2026.DOI
- [3]K. Huang, Q. Lan, Z. Liu, and L. Yang, “Semantic data sourcing for 6G edge intelligence,” IEEE Commun. Mag., vol. 61, no. 12, pp. 70–76, 2023.arxiv.org/abs/2301.00403
- [4]J. Shiraishi, M. Thorsager, S. R. Pandey, and P. Popovski, “TinyAirNet: TinyML model transmission for energy-efficient image retrieval from IoT devices,” IEEE Commun. Lett., vol. 28, no. 9, pp. 2101–2105, 2024.arxiv.org/abs/2311.04788
- [5]J. Shiraishi, S. R. Pandey, I. Leyva-Mayorga, and P. Popovski, “Wireless less memory approximation for energy-efficient task-specific IoT data retrieval,” in Proc. ICC 2026-IEEE Int. Conf. Commun., 2026.arxiv.org/abs/2510.26473
- [6]S. Liu, K. Patabiraman, T. Mescobroda, and B. G. Zorn, “Flikker: Saving DRAM refresh-power through critical data partitioning,” in Proc. sixteenth Int. Conf. Architectural support Program. Lang. operating Syst., vol. 1, pp. 213–224.
- [7]A. Raha, S. Sutar, H. Jayakumar, and V. Raghunathan, “Quality configurable approximate DRAM,” IEEE Trans. Comput., vol. 66, no. 7, pp. 1172–1187, 2016.DOI
- [8]D. T. Nguyen, N. H. Hung, H. Kim, and H.-J. Lee, “An approximate memory architecture for energy saving in deep learning applications,” IEEE Trans. Circuits Syst. I: Regular Papers, vol. 67, no. 5, pp. 1588–1601, 2020.DOI
- [9]A. Raha and V. Raghunathan, “Approximating beyond the processor: Exploring full-system energy-accuracy tradeoffs in a smart camera system,” IEEE Trans. Very Large Scale Integration (VLSI) Syst., vol. 26, no. 12, pp. 2884–2897, 2018.DOI
- [10]A. Das, S. K. Ghosh, A. Raha, and V. Raghunathan, “Towards energy-efficient collaborative inference using multi-system approximations,” IEEE Internet Things J., 2024.
- [11]P. Agheli, N. Pappas, P. Popovski, and M. Kountouris, “Effective communication: When to pull updates?” in ICC 2024-IEEE Int. Conf. Commun. IEEE, 2024, pp. 183–188.arxiv.org/abs/2311.06432
- [12]A. Kosta, N. Pappas, and V. Angelakis, “Age of information: A new concept, metric, and tool,” Found. Trends Netw., vol. 12, no. 3, pp. 162–259, 2017.
- [13]J. Deng et al., “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE Conf. comput. vision pattern recognit. IEEE, 2009, pp. 248–255.
- [14]J. Lin et al., “Mcunet: Tiny deep learning on iot devices,” Adv. Neural Inf. Process. Syst., vol. 33, pp. 11 711–11 722, 2020.
- [15]J. S. Rosenfeld, “Scaling laws for deep learning,” arXiv preprint arXiv:2108.07286, 2021.
- [16]F. Chiarotti et al., “Query age of information: Freshness in pull-based communication,” IEEE Trans. Commun., vol. 70, no. 3, pp. 1606–1622, 2022.
- [17]J. Shiraishi et al., “Content-based wake-up for energy-efficient and timely top-k IoT sensing data retrieval,” IEEE Trans. Commun., vol. 73, no. 12, pp. 13 983–13 998, 2025.arxiv.org/abs/2410.06026
- [18]B. Moons, K. Goetschalckx, N. Van Berckelaer, and M. Verhelst, “Minimum energy quantized neural networks,” in Proc. 2017 51st Asilomar Conf. Signals, Syst. Comput. IEEE, 2017, pp. 1921–1925.arxiv.org/abs/1711.00215
- [19]I. Bhati, Z. Chishti, S.-L. Lu, and B. Jacob, “Flexible auto-refresh: Enabling scalable and energy-efficient DRAM refresh reductions,” in Proc. 42nd Annu. Int. Symp. Comput. Architecture, 2015, pp. 235–246.
- [20]S. M. Z. Khan et al., “An empirical modeling for the baseline energy consumption of an NB-IoT radio transceiver,” IEEE Internet Things J., vol. 8, no. 19, pp. 14 756–14 772, 2021.
- [21]T. Besiroglu, E. Erdil, M. Barnett, and J. You, “Chinchilla scaling: A replication attempt,” arXiv preprint arXiv:2404.10102, 2024.arxiv.org/abs/2404.10102