Statistical Analysis for Energy-Efficient Satellite Edge Computing with Latency Guarantees Nicolai Dalsgaard Lyholm∗ , Beatriz Soret†∗ , Tijana Devaja∗ , Thomas Gundgaard Mulvad∗ , Cedomir Stefanovic∗ , and Israel Leyva-Mayorga∗ ∗ Department of Electronic Systems, Aalborg University, Denmark † Telecommunications Research Institute, University of Málaga, Spain
I. I NTRODUCTION LEO satellite constellations are rapidly transforming Earth Observation (EO) by enabling continuous capture of highresolution imagery at a large scale. However, the volume of data generated on-board far exceeds the available downlink capacity, creating a fundamental bottleneck in traditional store-and-forward architectures [1]. This disparity has motivated the shift towards Orbital Edge Computing (OEC), where computational resources are deployed and exploited directly on the satellite in order to process data closer to the point of acquisition, transmitting only the extracted semantic information to ground [2]. By performing tasks such as object detection, cloud filtering, and scene classification onboard, OEC dramatically reduces the required Downlink (DL) bandwidth enabling decision making for time-sensitive applications including maritime surveillance, disaster response, and environmental monitoring. A key enabler of OEC is the advancements and increased availability of GPU-accelerated edge hardware capable of executing Artificial Intelligence (AI) algorithms within the strict size, weight, and power constraints of satellite platforms. Recent work has shown that such devices exhibit complex,
Relay Sat
Destination Sat Semantic extraction
Serving Sat ues
t
GPU
Tproc
Altitude h
req
Abstract—Being able to provide latency guarantees for orbital edge computing applications through Low Earth Orbit (LEO) satellite constellations is a major milestone for their integration into 5G and 6G networks. However, achieving this is fundamentally challenged by the inherent randomness in both communication and computing latency, driven by complex network dynamics, satellite motion, and hardware variability. In this paper, we perform a statistical analysis of the latency of satellite edge computing using representative computing hardware and an object detection algorithm running on a satellite image dataset. The resulting model captures the tradeoff between data availability and estimation uncertainty, enabling data-driven optimization methods to meet latency targets with statistical guarantees while minimizing energy consumption. Our results show that parametric estimation and quantile regression for the execution time of the image processing algorithms can be effectively combined with models for the communication latency to select an optimal GPU clock frequency. This achieves a 95% probability of meeting a 500 ms end-to-end deadline while reducing energy consumption by more than 50% compared to a baseline that relies on a Chebyshev–Cantelli inequality to bound execution-time quantiles. The proposed framework is generalizable across satellite edge computing workloads and hardware platforms.
Use r
arXiv:2605.10215v1 [cs.NI] 11 May 2026
Emails: {ndly, bsa, tde, tgm, cs, ilm}@es.aau.dk,
Earth Area of interest
User (UE)
Requests image data
End-to-end deadline budget TUL
TISL
Tproc
On-board GPU inference
TISL
TDL
TE2E (deadline)
Fig. 1. Satellite edge computing scenario and end-to-end latency budget TE2E , decomposed into uplink (TUL ), inter-satellite (TISL ), on-board processing (Tproc ) and downlink (TDL ) components.
workload-dependent performance characteristics: execution times vary with operating frequency, power mode and the computational demands of the inference task [3], [4]. Understanding and accurately modeling these characteristics is essential for scheduling and resource management in orbital environments, where the energy budgets are tightly constrained and processing must be completed within strict timing windows dictated by the orbital geometry and the mission planning. The relevance of GPU-accelerated hardware in space is further underscored by NVIDIA’s announcement at GTC 2026 that the Jetson Orin platform is being adopted for on-orbit AI inference by satellite operators including Kepler communication and Sophia Space [5], establishing Jetson Orin as a representative platform for OEC research. In this paper, we consider a satellite edge computing scenario where users on ground request information contained in satellite images from a specific geographic area. The users send requests to their serving satellite, which relays the information to a destination satellite with visibility over the area of interest. The destination satellite captures the images and, based on our statistical model, chooses the best processing frequency to extract the semantic information from the image with minimum energy consumption while achieving statistical guarantees of meeting the latency constraints of the application. Our main contributions are:
1) We present a statistical framework to characterize the per-image execution time of an object detection algorithm on representative on-board GPU hardware as a Gamma-distributed random variable with frequencydependent shape and scale parameters. 2) We formulate the selection of the GPU clock frequency as an energy-minimization problem under a quantilebased reliability constraint on the end-to-end latency, incorporating a finite-blocklength model of the satellite to user request transmission. 3) We compare the performance of our framework against a benchmark that obtains a closed-form upper bound on the optimal frequency from the Chebyshev-Cantelli inequality without imposing a distributional assumption. 4) We quantify the impact of sample size on the ability of the parametric model to yield reliable scheduling decisions, highlighting the tradeoff between model training effort and estimation uncertainty. II. S YSTEM MODEL We consider the 5G Non-Terrestrial Network (NTN) scenario illustrated in Fig. 1, in which a user (UE) on ground requests information contained in satellite images from a specific geographical area. The request is transmitted to the serving satellite via the Uplink (UL), forwarded through InterSatellite Links (ISLs) to a destination satellite with visibility of the target area. The destination satellite captures Nimg images, processes them to extract semantic information (e.g. ship positions via object detection), and returns the result to the user through ISLs and the DL. Both the request and the inference output are small data blocks exchanged over the network, while the computational load is concentrated onboard the satellite. The task is latency-constrained: the End-to-End (E2E) delivery must complete within a maximum latency TE2E . The satellite chooses an operating frequency for its on-board GPU computing hardware, which minimizes energy consumption while statistically fulfilling the latency constraint.
Signal-to-Noise Ratio (SNR) for communication between the UE and the satellite, separated by a distance d is γℓ (d) =
Pℓ Gsat GUE , L(d) φ σℓ2 Ψ
where φ is the pointing loss, σℓ2 is the noise power at the receiver and Ψ is a log-normal shadow fading variable, i.e. 10 log10 (Ψ) ∼ N (0, σsf2 ). Given that the request for satellite image data requires only a few bytes and the transmission power at the UEs is limited, we consider Finite Blocklength (FBL) effects in the UL, where the block error probability depends on the block length n, code rate R, and instantaneous SNR γ [6], [7] as r n C(γ) − R , (3) ϵ(γ) = Q V (γ) with C(γ) = log2 (1 + γ) being the Shannon capacity and 2 γ+2 V (γ) = γ (1+γ) 2 log2 (e) being the channel dispersion. An Automatic Repeat Request (ARQ) mechanism is implemented to retransmit failed packets and, hence, provide reliable data delivery despite finite blocklength effects. Due to a considerably higher transmit power at the satellite PDL ≫ PUL , NACK transmissions are assumed to be error free and the number of transmission attempts follows a geometric distribution with success probability 1 − ϵ(γ). The uplink blocklength n (in channel uses) required by the finite blocklength model is mapped onto the orthogonal frequency-division multiplexing (OFDM) resource grid of 5G NTN, where NSC channel uses can be conveyed simultaneously within each symbol duration Tsymb . Therefore, the UL delay per transmission attempt, including propagation, is Ttx = Tsymb ⌈n/NSC ⌉ + d/vc . Hence, the uplink delay distribution including retransmissions and NACK delays is Pr TUL (R, d) = Ttx + x(Ttx + TNACK ) = (1 − ϵ(γUL (d)))ϵ(γUL (d))x ,
where vc is the speed of light and d is the distance between the user and the serving satellite. Note that the latter is dependent on the altitude of the serving satellite h and its elevation angle at the specific time instant. Next, let Gsat and GUE denote the antenna gain of the satellite and the UE, respectively, and let Pℓ be the transmission power for link ℓ ∈ {DL, UL}. The
(4)
where x is the number of failed attempts and TNACK is the fixed NACK delay. The expected uplink delay is
A. Communication Model The satellites communicate with each other through FreeSpace Optical (FSO) ISLs that operate at a fixed rate Risl bps. UL and DL communication between the user and the serving satellite occurs over an additive white Gaussian noise (AWGN) channel using radio-frequency links with frequency fcomm Hz. At a given time instant, the satellite-to-user links experience free-space path loss 2 4πd fcomm , (1) L(d) = vc
(2)
E[TUL (R, d)] = Ttx +
(Ttx + TNACK ) ϵ(γUL (d)) 1 − ϵ(γUL (d))
(5)
While the use of short block lengths may result in FBL effects in the DL, the error probability is negligible due to the significantly higher transmission power at the satellites, assuming the same block length n as in the UL. Hence, TDL = Tsymb ⌈n/NSC ⌉ + dDL /vc is treated as deterministic. Retransmissions are also neglected during inter-satellite communication, such that the delay at each ISL is composed of a transmission and a propagation component. For a block of size n bits transmitted over a path P, where the path itself determines the number of hops, and where the inter-satellite distance of each link e ∈ P is denoted as dISL (e), the roundtrip ISL latency is given by X n 2 dISL (e) ISL + TISL = 2 Tsymb , (6) ISL vc NSC e∈P
ISL ISL where Tsymb and NSC are the duration of a symbol and the number of subcarriers in the ISL, respectively.
B. Processing Model Each processing node in the constellation is characterized by: the number of processing cores Ncores , the maximum operating frequency fmax , the Floating Point Operations (FLOPs) per clock cycle per core NFLOPs , and the maximum power consumption Pmax . The power consumption at operating frequency fp follows the cubic scaling model 3 fp . (7) P (fp ) = Pmax fmax The execution time for processing a single image at frequency fp is modeled by the Bulk-Synchronous Parallel (BSP) model as CW Texec (fp ) = + Tsync , (8) Ncores NFLOPs fp where W is the algorithmic work in FLOPs, C ≥ 1 is a Random Variable (RV) capturing processing inefficiency, and Tsync captures synchronization overhead within the processing unit.
and BIC statistics among the candidate distributions. These findings are in line with observations in the literature [8], [9]. Building on this, the frequency dependent shape and scale parameters are estimated using polynomial regression α̂(fp ) =
(9)
where µC = E[C] and µsync = E[Tsync ]. The least squares estimators for µC and µsync for the case where a sample of size Ns for Texec (fp ) is available for multiple values of fp are, respectively Ns Ns µ̂C X 1 1 X Ti − µ̂sync = Ns i=1 Ns i=1 fpi ! Ns Ns Ns X X Ti 1 X 1 − Ti j i Ns i=1 f i=1 p j=1 fp µ̂C = , !2 N Ns s X 1 1 1 X − i 2 Ns i=1 fpi i=1 f
(10)
(11)
p
where Ti is the execution time (i.e., outcome) of the i-th image (i.e., experiment) using processing frequency fpi . Beyond the mean, the full distribution of the execution time is essential for providing scheduling guarantees. Throughout this paper, we model the per-image execution time for algorithms running on GPU hardware as a Gamma-distributed RV Texec (fp ) ∼ Gamma(α(fp ), θ(fp )),
θ̂(fp ) =
L X
cθ,i fpi ,
(13)
i=0
where L is the polynomial order and cα,i and cθ,i are the fitted coefficients for the shape and scale respectively. When processing Nimg images sequentially at a constant operating frequency fp , the total processing time is the sum of Nimg Independent and Identically Distributed (i.i.d.) Gamma RVs. Using the Moment Generating Function (MGF), we have N
img MTexec (t) = Πi=1 (1 − θ(fp )t)
−α(fp )
−Nimg α(fp )
= (1 − θ(fp )t)
,
(14)
which is the MGF of a Gamma distribution with shape Nimg α(fp ) and scale θ(fp ). Therefore, Nimg
Texec (fp , Nimg ) =
X
Texec,i (fp )
i=1
∼ Gamma (Nimg α(fp ), θ(fp )) . (15)
Taking the expectation of the BSP model (8), the mean execution time is given as µC W + µsync , Ncores NFLOPs fp
cα,i fpi ,
i=0
C. Statistical Characterization of Execution Time
E[Texec (fp )] =
L X
(12)
with mean E[Texec (fp )] = α(fp ) θ(fp ) and variance Var(Texec (fp )) = α(fp ) θ(fp )2 . The goodness of fit of the Gamma distribution to the execution time with the considered hardware and algorithm was the highest among a wide range of candidate distributions, reaching a Kolmogorov-Smirnov statistic of 0.05, further supported by achieving the best AIC
D. Optimization The E2E latency consists of the communication times for the UL, the ISL, and the DL, and the processing time. Once the data reaches the satellite responsible for computation, the communication delays, originally modeled as random variables, are realized as deterministic values. Then, the available processing time budget is written as Tproc = TE2E − TUL − TISL − TDL ,
(16)
where TE2E is the total E2E time budget. Note that, since P (fp ) is increasing monotonically, minimizing the operating frequency is equivalent to minimizing the energy consumption. Therefore, the optimal processing frequency for a latency-constrained application that requires a reliability ρth is the minimum operating frequency such that the processing completes within Tproc with a probability of at least ρth . Namely, for a given Nimg α(fp ), θ(fp ), the optimal processing frequency is fp∗ ≜ min
fp ∈ [fmin , fmax ]
s.t. FGamma Tproc ; Nimg α(fp ), θ(fp ) ≥ ρth .
(17)
E. Benchmark As a benchmark that uses only the first two moments of the execution time distribution, the Chebyshev-Cantelli inequality provides a one-sided bound on the probability of the total processing time exceeding the deadline, given by Pr(Texec (fp , Nimg ) ≥ t) ≤
Nimg σ 2 (fp ) Nimg σ 2 (fp ) + t − Nimg E[Texec (fp )]
2 .
(18)
(a) Jetson Nano
(b) Jetson AGX Orin
0.8
Probability Density
Probability Density
0.8 0.6 0.4 Fitted Distribution 306 MHz 714 MHz 1020 MHz
0.2 0.0
60
80
100
120
Execution Time (ms)
140
160
0.6 0.4 Fitted Distribution 306 MHz 816 MHz 1300 MHz
0.2 0.0
30
40
50
60
70
Execution Time (ms)
80
90
Fig. 2. Empirical execution time distribution for (a) NVIDIA Jetson Nano and (b) Jetson AGX Orin at selected GPU operating frequencies. The solid black lines represent the fitted Gamma distributions.
valid for t > Nimg E[Texec (fp , Nimg )], where σ 2 (fp ) = Var(Texec (fp , Nimg )). Imposing the reliability requirement that the right-hand side is at most 1 − ρth and substituting the BSP mean from (9) yields an upper bound on the optimal operating frequency Nimg µC W Ncores NFLOPs 1 p . · t − Nimg µsync − σ(fp ) Nimg ρth /(1 − ρth )
fp∗ ≤ fpupper (Nimg ) ≜
(19)
F. Learning the Computation Time Distribution The optimization in eq. (17) requires knowledge of the Gamma parameters α(fp ) and θ(fp ), which must be learned empirically. In practice, a newly deployed satellite cannot afford to conduct an exhaustive sampling campaign before becoming operational. We study how scheduling reliability degrades when the Gamma parameters are estimated from a limited sample size. Let S denote the full set of maritime images in the dataset. To emulate the situation in which only a limited number of execution time observations are available for parameter fitting, we draw a subset Sk of size |Sk | = Ns from S with replacement. Sampling with replacement allows Ns > |S|, with repeated draws of the same image yielding distinct samples that capture the hardware’s variability between executions. Each image in Sk is executed once per available operating frequency, yielding Ns execution time samples at each frequency. From these samples the Gamma parameters are estimated via Maximum Likelihood Estimation (MLE), producing a subset characterization (α̂k (fp ), θ̂k (fp )). To capture the variability of this characterization across different sample realizations, we repeat the procedure K times for each sample size Ns , generating K independent subset characterizations and conse∗ quently, K frequency selections fˆp,k obtained by solving (17) ∗ under each subset. Each estimated frequency fˆp,k is evaluated
against the ground-truth Gamma distribution to compute the probability of missing the deadline due to a poor estimate (k) ∗ ∗ Pmiss = 1 − FTexec (Tproc ; Nimg α(fˆp,k ), θ(fˆp,k )).
(20)
(k)
If Pmiss ≈ 1 − ρth , the subset characterization is reliable. On (k) the other hand, if Pmiss ≫ 1−ρth the subset underestimated the execution time tail and operates with false confidence. Vice (k) versa, if Pmiss ≪ 1 − ρth , the subset is overly conservative, selecting a too high operating frequency and wasting energy. III. R ESULTS We evaluate the proposed framework in two parts. First, we validate the Gamma distribution model and assess its increase in accuracy as the sample size increases. Then, we compare the energy consumption of the Chebyshev-Cantelli baseline, imposing no distributional assumption on the execution time, against the Gamma quantile-based scheduling under varying workload and link conditions. A. Experimental Setup We validate the proposed framework using YOLOv8m [10] for ship detection in satellite imagery, executed on two NVIDIA Jetson platforms: the Orin Nano and the AGX Orin. While the framework is algorithm- and platform agnostic, this combination reflects two tiers of realistic OEC deployment scenarios [5]. We use the Ships-Google-Earth dataset, which is representative of the target Earth observation application and contains 56 satellite images of maritime scenes with varying ship densities [11]. For the ground truth characterization, each image is executed 1000 times at every GPU frequency. The Gamma parameters α(fp ), θ(fp ) are fitted via MLE and their frequency dependence is modeled by third-order polynomial regression. Communication, computing and simulation parameters are summarized in Table I. Throughout the evaluation, we set the reliability requirement ρth = 0.95 and the E2E deadline to TE2E = 500 ms unless stated otherwise.
0.7
60
Jetson AGX Jetson Nano Target 0.05
0.6
Nano infeasible Nimg ≥ 8
50
0.5
AGX infeasible Nimg ≥ 14
40
E (J)
Pmiss
0.4 0.3 0.2
30
20
0.1
AGX, Gamma quantile AGX, Cantelli bound Nano, Gamma quantile Nano, Cantelli bound
10
0.0 10
1
10
2
10
3
10
4
Ns (sample size)
Fig. 3. Deadline miss probability Pmiss versus sample size Ns . Shaded regions show variation across the K subsets for any choice of Ns .
0
0
5
10
15
20
25
30
Nimg
Fig. 4. Processing energy E versus number of requested images Nimg for the proposed Gamma quantile-based scheduling and the Chebyshev-Cantelli baseline, under error-free communication.
B. Execution Time Characterization Fig. 2 shows the empirical execution time distributions at three selected frequencies, overlaid with the fitted Gamma Probability Density Functions (pdfs). The Gamma model captures the shape, location and spread of the execution time samples across the full frequency range, confirming its suitability for characterizing inference execution times on GPU edge hardware. The frequency-dependent parameters α(fp ) and θ(fp ) are well described by the polynomial model, achieving a coefficient of determination > 0.99 for α and > 0.95 for θ on both platforms. A key practical question is the sample size needed to learn a reliable characterization of the execution time. Fig 3 evaluates (k) this by plotting the deadline miss probability Pmiss as a function of the sample size Ns used to fit the Gamma parameters. For each Ns , K = 100 subsets are drawn with replacement, each producing an independent characterization and frequency selection via (17). After this, the miss probability is evaluated against the ground truth distribution. At small Ns , the subset characterizations systematically underestimate the execution time tail, leading to too low frequency selections and miss probabilities significantly above the 1 − ρth = 0.05 target. As Ns increases, the mean converges towards the target and the spread across subsets narrows. Beyond Ns = 1000, the miss probability stabilizes within a small margin of the design target for both platforms, indicating that reliable scheduling can be achieved with a modest number of executions of the algorithm. C. E2E Optimization Fig 4 compares the processing energy consumption resulting from Gamma quantile-based scheduling against the Chebyshev-Cantelli baseline as a function of the number of requested images Nimg . We focus on the impact of processing time on the scheduling decision. To this end, the evaluation is conducted under conditions where ϵ(γUL ) ≈ 0, corresponding
to high elevation angles, such that uplink ARQ retransmissions are negligible. As Nimg increases, the image processing time budget Tproc shrinks, requiring higher operating frequencies and thus greater energy consumption. The Gamma quantile-based method consistently selects a lower operating frequency than the Cantelli bound, since it exploits the full distributional knowledge, whereas the Cantelli bound is limited to the first two moments. The gap between the two methods increases as Nimg grows and the deadline becomes tighter, highlighting the importance of distribution-aware scheduling under stringent deadlines. Beyond Nimg = 8 for the Nano and Nimg = 14 for the AGX, even the maximum operating frequency cannot satisfy the reliability constraint and the deadline becomes infeasible. In practice, the available processing budget Tproc depends on the communication delays, which in turn depend on the satellite to user geometry. At low elevation angles the increased path-loss raises the UL error probability ϵ(γUL ), leading to more ARQ retransmissions and a larger expected UL delay, thereby further reducing the time budget available for processing. Fig 5, evaluates this effect for both platforms at selected feasible values of Nimg , sweeping the elevation angle from near-zenith to near-horizon. At high elevation angles, few ARQ retransmissions occur and the energy consumption is dominated by the processing workload, matching the behavior observed in Fig 4. As the elevation angle and γUL decrease, the UL error probability ϵ(γUL ) increases, causing more ARQ retransmission and a larger expected TUL . This reduces the processing budget, forcing the satellite to increase its operating frequency. Below a critical elevation angle (indicated by the shaded region), the expected UL delay consumes the entire E2E deadline, making the task infeasible regardless of the processing strategy.
TABLE I S IMULATION PARAMETERS
25
[TUL ] ≥ TE2E
20
E (J)
15
10
AGX, Nimg = 7 AGX, Nimg = 10 AGX, Nimg = 12
Nano, Nimg = 3 Nano, Nimg = 5 Nano, Nimg = 6
5
0 2
3
5
7
10
15
30
90
Elevation angle [deg]
Fig. 5. Processing energy E versus satellite elevation angle for selected feasible values of Nimg on both platforms. The shaded region marks elevations at which E[TUL ] ≥ TE2E , such that the task becomes infeasible.
IV. C ONCLUSION This paper presented a statistical framework that enables the provision of reliability guarantees on the end-to-end latency of satellite edge computing applications while minimizing energy consumption. Our framework was validated on an Earth observation application with satellite image dataset and the YOLOv8m object detection algorithm executed on the NVIDIA Jetson Orin Nano and AGX platforms. In this scenario, the execution time was characterized as a Gammadistributed random variable with frequency-dependent shape and scale parameters. Our method reduces the processing energy consumption by more than 50% when compared to the selected benchmark, based on the Chebyshev–Cantelli inequality, while meeting a 95% reliability target on a 500 ms end-to-end deadline, with the gap widening as the deadline tightens. Furthermore, we showed that reliable parameter estimates can be obtained from a relatively small sample size of approximately 1000, making the characterization practical to learn on-board. Future work will focus on the extension and integration of our statistical framework into a digital twin (DT) framework to enable continuous model adaptation and constellation-wide multi-service integration and orchestration. V. ACKNOWLEDGEMENT This work was supported, in part, by the Velux Foundation, Denmark, through the Villum Investigator Grant WATER, nr. 37793 and by Danmarks Frie Forskningsfond (DFF) Project “3D-Twin” under Grant 10.46540/4264-00153B. The work of B. Soret is partially supported by the Spanish Ministerio de Ciencia e Innovación under grant PID2022-136269OB-I00. R EFERENCES [1] Z. Yin, C. Wu, C. Guo, Y. Li, M. Xu, W. Gao, and C. Chi, “A comprehensive survey of orbital edge computing: Systems, applications, and algorithms,” Chinese Journal of Aeronautics, vol. 38, no. 7, p. 103316, Jul. 2025.
Parameter
Symbol
Setting
Communication Carrier frequency [GHz] Bandwidth [kHz] Satellite antenna gain [dBi] UE antenna gain [dBi] Total number of satellites Number of ISL hops Altitude of deployment [km] Downlink transmission power [W] Uplink transmission power [W] Pointing loss [dB] Noise spectral density [dBm/Hz] Shadow fading std. [dB] Number of subcarriers Number of OFDM symbols Subcarrier spacing [Hz] Number of slots allocated Payload size [bits] Number of channel uses Transmission rate [bpcu]
fcomm B Gsat GUE Nsats Nhops h PDL PUL φdB N0,dB σsf Nsc Nsym SCS Nslots D n R
2 180 30 0 12 4 600 75 0.2 0.3 −176.31 4 12 14 15e3 4 1500 672 2.23
Computing with NVIDIA Jetson Orin Number of cores [W] Ncores Max. power consumption [W] Pmax Max. clock frequency [GHz] fmax Number of FLOPs per Hz NFLOPs Processing inefficiency µC Mean synchronization time [ms] µsync h i Mean execution time (p) E Texec (fmax ) at max. frequency [ms]
Nano 1024 25 1.02 2 1.071 17.48
AGX 2048 60 1.3 2 1.122 14.14
61.19
32.63
[2] B. Denby and B. Lucia, “Orbital Edge Computing: Machine Inference in Space,” IEEE Computer Architecture Letters, vol. 18, no. 1, pp. 59– 62, Jan. 2019. [3] P. S.K., V. Hegde, K. Patchava, A. Das, and Y. Simmhan, “Performance Characterization of Containerized DNN Training and Inference on Edge Accelerators,” in Proc. 30th International Conference on High Performance Computing, Data, and Analytics (HiPC), Dec. 2023. [4] B. Soret, A. M. Mercado-Martínez, A. Jurado-Navas, N. D. Lyholm, M. Moretti, P. Popovski, and I. Leyva-Mayorga, “Edge Intelligence for Satellite-based Earth Observation: Scheduling Image Acquisition and Processing,” Apr. 2026, arXiv:2604.05937. [5] NVIDIA, “NVIDIA launches space computing, rocketing AI into orbit,” Press Release, Mar. 2026, accessed: 2026-04-08. [Online]. Available: https://nvidianews.nvidia.com/news/space-computing [6] Y. Polyanskiy, H. V. Poor, and S. Verdu, “Channel coding rate in the finite blocklength regime,” IEEE Transactions on Information Theory, vol. 56, no. 5, pp. 2307–2359, 2010. [7] T. Devaja, M. Petkovic, C. Wang, M. Beko, and D. Vukobratovic, “On error probability analysis of short-packet communications in massive Internet of Things,” IEEE Access, vol. 12, pp. 67 107–67 116, 2024. [8] J. R. Lorch and A. J. Smith, “Improving dynamic voltage scaling algorithms with PACE,” in Proc. ACM International Conference on Measurement and Modeling of Computer Systems (SIGMETRICS), New York, NY, USA, 2001, p. 50–61. [9] S. Suman, C. Stefanovic, S. Došen, and P. Popovski, “Analysis and Optimization of the Latency Budget in Wireless Systems with Mobile Edge Computing,” in Proc. IEEE International Conference on Communications, May 2022, pp. 5029–5034. [10] G. Jocher, A. Chaurasia, and J. Qiu, “Ultralytics YOLOv8,” 2023, accessed: 2026-02-20. [Online]. Available: https://github.com/ultralytics/ultralytics [11] K. Stavrakakis, “Ships-Google-Earth Dataset,” https://universe.roboflow.com/k–stavrakakis/ships-google-earth, May 2022, accessed: 2026-02-20.