Paper deep dive
Joint Optimization of Memory and Computing Frequency for Energy-Efficient DNN Inference
Yunchu Han, Zhaojun Nan, Sheng Zhou, Zhisheng Niu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 8/17/2026, 4:28:52 AM
Summary
This paper addresses the energy efficiency of Deep Neural Network (DNN) inference on mobile devices by jointly optimizing memory frequency, computing frequency, and communication resources (transmission power and bandwidth). Unlike previous studies that primarily focus on computing frequency via Dynamic Voltage and Frequency Scaling (DVFS), this work highlights the significant impact of memory frequency on inference latency and energy consumption. The authors formulate a mixed-integer nonlinear programming (MINLP) problem to minimize total energy consumption under deadline constraints. They derive near-optimal closed-form solutions for local inference using convex optimization and optimal closed-form solutions for edge inference transmission power. A low-complexity heuristic algorithm is proposed to solve the overall problem. Simulation results demonstrate that the proposed approach reduces energy consumption by up to 10.4% compared to existing methods and achieves performance within 2.5% of the optimal solution for local inference.
Entities (10)
Relation Signals (7)
ResNet152 → deployedon → Jetson TX1
confidence 96% · we deploy ResNet152 [14] and VGG19 [15] on Jetson TX1
VGG19 → deployedon → Jetson TX1
confidence 96% · we deploy ResNet152 [14] and VGG19 [15] on Jetson TX1
Memory Frequency → impacts → Energy Consumption
confidence 95% · 80% and 92% reductions can be achieved for the average energy consumption by only adjusting the memory frequency
Memory Frequency → impacts → DNN Inference Time
confidence 95% · by increasing the memory frequency from 0.1 GHz to 1.6 GHz, the average inference time can be reduced by 84% and 93%
DVFS → adjusts → Computing Frequency
confidence 94% · most existing studies focus on dynamic voltage and frequency scaling (DVFS) for adjusting the computing frequency
Joint Optimization → reduces → Energy Consumption
confidence 94% · our proposed algorithm significantly reduces the energy consumption of devices by up to 10.4% compared to other methods
Heuristic Algorithm → solves → MINLP Problem
confidence 93% · we propose a low-complexity heuristic algorithm to effectively solve the overall problem with polynomial time complexity
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Deep neural network (DNN) inference on mobile devices often incurs high latency and energy consumption due to limited computing and memory resources. To enable energy-efficient DNN inference, most existing studies focus on dynamic voltage and frequency scaling (DVFS) for adjusting the computing frequency, while the impact of memory frequency on the inference performance has been greatly overlooked. In this paper, we consider the impact of memory frequency and computing frequency on DNN inference time, and jointly optimize these two frequencies together with communication resources for energy-efficient DNN inference. Based on a realistic inference time model, we formulate an optimization problem to minimize the energy consumption of all mobile devices under the deadline constraint. For local inference, we derive a near-optimal closed-form solution via convex optimization, while an optimal closed-form solution for transmission power is obtained for edge inference with the given bandwidth. Furthermore, we propose a low-complexity heuristic algorithm to effectively solve the overall problem with polynomial time complexity. Simulation results based on measured data show that the proposed near-optimal solution for local inference can achieve optimal performance under strict deadline constraints, with a performance gap of up to 2.5% compared with the optimal solution. Meanwhile, our proposed algorithm significantly reduces the energy consumption of devices by up to 10.4% compared to other methods.
Tags
Links
- Source: https://arxiv.org/abs/2608.13863v1
- Canonical: https://arxiv.org/abs/2608.13863v1
Trouble viewing inline? Open PDF directly →
Full Text
33,566 characters extracted from source content.
Expand or collapse full text
Joint Optimization of Memory and Computing Frequency for Energy-Efficient DNN Inference Yunchu Han1, Zhaojun Nan2, Sheng Zhou1, and Zhisheng Niu1 Affiliation: 1Beijing National Research Center for Information Science and Technology Department of Electronic Engineering, Tsinghua University, Beijing 100084, China Affiliation: 2School of Electronics and Internet of Things Chongqing Polytechnic University of Electronic Technology, Chongqing 401331, China Affiliation: Emails: hyc23@mails.tsinghua.edu.cn, nanzhaojun@cquet.edu.cn, sheng.zhou@, niuzhs@tsinghua.edu.cn Abstract Deep neural network (DNN) inference on mobile devices often incurs high latency and energy consumption due to limited computing and memory resources. To enable energy-efficient DNN inference, most existing studies focus on dynamic voltage and frequency scaling (DVFS) for adjusting the computing frequency, while the impact of memory frequency on the inference performance has been greatly overlooked. In this paper, we consider the impact of memory frequency and computing frequency on DNN inference time, and jointly optimize these two frequencies together with communication resources for energy-efficient DNN inference. Based on a realistic inference time model, we formulate an optimization problem to minimize the energy consumption of all mobile devices under the deadline constraint. For local inference, we derive a near-optimal closed-form solution via convex optimization, while an optimal closed-form solution for transmission power is obtained for edge inference with the given bandwidth. Furthermore, we propose a low-complexity heuristic algorithm to effectively solve the overall problem with polynomial time complexity. Simulation results based on measured data show that the proposed near-optimal solution for local inference can achieve optimal performance under strict deadline constraints, with a performance gap of up to 2.5%2.5\% compared with the optimal solution. Meanwhile, our proposed algorithm significantly reduces the energy consumption of devices by up to 10.4%10.4\% compared to other methods. I Introduction The rapid development of deep neural networks (DNNs) has significantly advanced computer vision and artificial intelligence applications [1], [2]. However, the intensive computation and memory access requirements of DNN inference impose critical challenges for mobile devices with limited energy and computing resources. To enable low-latency and energy-efficient DNN inference, mobile edge computing (MEC) [3] has emerged as a promising technology, allowing computation tasks to be offloaded from mobile devices to nearby edge servers. Furthermore, edge intelligence [4] has been proposed as a new paradigm that enables real-time and intelligent services at the wireless network edge (e.g., base stations and road side units). However, mobile devices still need to manage their local computing and memory resources efficiently to balance inference latency and energy consumption. To address this issue, the dynamic voltage and frequency scaling (DVFS) [5], [6] technique has been widely studied to achieve a tradeoff between latency and energy consumption by dynamically adjusting the operating frequency of processors. Existing DVFS-based approaches mainly focus on the adjustment of computing frequency (e.g., CPU frequency, GPU frequency) to reduce latency or energy consumption. For example, the authors in [7] design a robust DNN partitioning and resource allocation algorithm to handle uncertain inference time, where DVFS is adopted to adjust the computing frequency for local inference. In addition, DVFS can be combined with the batch processing technique to reduce the energy consumption [8] or increase the throughput [9]. Fig. 1: Impact of memory frequency scaling on DNN inference time and energy consumption for ResNet152 and VGG19 on Jetson TX1. However, recent studies have revealed that memory frequency scaling also plays a significant role in the overall latency, energy and accuracy characteristics of DNN inference, especially for memory-intensive DNN models [10], [11], [12], [13]. In [10], the impact of joint memory frequency and computing frequency scaling on DNN inference time for edge devices is characterized and analyzed. In [11], the impact of memory frequency and computing frequency on energy efficiency of DNN inference is investigated, respectively, and a reinforcement learning algorithm is proposed to optimize these frequencies. In [12], the authors mainly analyze the impact of memory frequency and computing frequency on kernel-based applications. In [13], the impact of memory voltage scaling on accuracy and resilience of DNNs for edge devices is presented. To further investigate the impact of memory frequency scaling on latency and energy consumption of DNN inference, we deploy ResNet152 [14] and VGG19 [15] on Jetson TX1 to evaluate the corresponding performance, as shown in Fig. 1. It is observed that by increasing the memory frequency from 0.10.1 GHz to 1.61.6 GHz, the average inference time can be reduced by 84%84\% and 93%93\% for ResNet152 and VGG19, respectively. Meanwhile, 80%80\% and 92%92\% reductions can be achieved for the average energy consumption by only adjusting the memory frequency. Therefore, memory frequency scaling can also significantly impact the inference time and energy consumption. Despite these potentials, the joint optimization of memory frequency and computing frequency for energy-efficient DNN inference has not been fully studied. In this paper, we consider an edge intelligence system that adopts the joint optimization of memory frequency, computing frequency and communication resources to reduce the energy consumption of mobile devices. Specifically, we derive an optimal solution to the local inference problem in some special cases. Moreover, we analyze the upper bound of the corresponding problem and derive a near-optimal closed-form solution. Meanwhile, the analysis for the edge inference problem provides an optimal closed-form solution for the transmission power, and a heuristic algorithm is proposed to efficiently solve the overall optimization problem. Simulation results show that the proposed near-optimal solution achieves a performance within 2.5%2.5\% of the optimal result for local inference, and reduces the energy consumption by up to 10.4%10.4\% compared with other methods, demonstrating the effectiveness of the proposed approach. I System Overview I-A System Model As shown in Fig. 2, we consider an edge intelligence system consisting of N mobile devices and an edge server. The set of all mobile devices is denoted as ≜1,2,…,NN \1,2,…,N\. The orthogonal frequency division multiple access (OFDMA) technology is adopted to reduce the interference among mobile devices, and we assume that the bandwidth allocated to each device does not overlap. Each device needs to execute DNN inference tasks within the given deadline DnD_n. We consider the binary offloading mode (i.e., xn∈0,1x_n∈\0,1\) in this work, where mobile device n can either execute local inference (i.e., xn=1x_n=1) or edge inference by offloading the inference task (i.e., xn=0x_n=0). For local inference, mobile devices can adjust the memory frequency and computing frequency to save the energy consumption of DNN inference with the deadline constraint. For edge inference, mobile devices can change the transmission power to improve the transmission energy while meeting the communication latency constraint. The edge inference time is negligible due to the powerful computing capacity of the MEC server, and the energy consumption of edge inference is ignored here because the MEC server is usually powered by the grid [16]. In addition, the latency of downloading inference results is ignored due to their small data volume [17]. Fig. 2: Illustration of an edge intelligence system that jointly optimizes memory and computing frequencies, transmission power and bandwidth. I-B Inference Time and Energy Consumption Model We perform real-world experiments to obtain the average inference time under different combinations of memory frequencies and computing frequencies [10], and the impact of memory frequency fn,memf_n,mem and computing frequency fn,comf_n,com on DNN inference time is formulated as tnloc=anfn,mem−bn+cnfn,com−dn,∀n∈,t_n^loc=a_nf_n,mem^-b_n+c_nf_n,com^-d_n,∀ n , (1) The dynamic power consumption of CMOS circuit is denoted by P=αCV2fP=α CV^2f, where α, C, V and f denote the activity factor, the capacity, the supply voltage and the frequency, respectively [18]. Since the power consumption consists of the memory and computing power [19], the power consumption of mobile device n for executing inference is written as pnloc=κn,memfn,mem3+κn,comfn,com3+σn,∀n∈, p_n^loc= _n, memf_n, mem^3+ _n, comf_n, com^3+ _n,∀ n , (2) where the details are given in [10]. Therefore, the energy consumption of mobile device n for executing inference is enloc=pnloctnloc,∀n∈. e_n^loc=p_n^loct_n^loc,∀ n . (3) I-C Transmission Time and Energy Consumption Model The transmission rate of mobile device n can be written as rn=Bnlog2(1+pnhnσ2),∀n∈, r_n=B_n _2 (1+ p_nh_nσ^2 ),∀ n , (4) where BnB_n, pnp_n, hnh_n and σ2σ^2 denote the allocated bandwidth, transmission power, channel gain and noise power, respectively. Then, the transmission latency for task offloading of mobile device n can be written as tntran=snrn,∀n∈, t_n^tran= s_nr_n,∀ n , (5) where sns_n is the data volume of the computation task. The corresponding energy consumption of transmission is entran=pntntran,∀n∈. e_n^tran=p_nt_n^tran,∀ n . (6) I Problem Formulation and Solutions We focus on minimizing the sum of the energy consumption of all mobile devices by optimizing the binary offloading decision xnx_n, transmission power pnp_n, bandwidth BnB_n, memory frequency fn,memf_n,mem and computing frequency fn,comf_n,com, while the deadline constraint should be satisfied. The optimization problem can be formulated as P1:minxn,pn,Bn,fn,mem,fn,com∑n∈(xnenloc+(1−xn)entran) 1: _x_n,p_n,B_n,f_n,mem,f_n,com _n (x_ne_n^loc+ (1-x_n )e_n^tran ) (7a) s.t.xntnloc+(1−xn)tntran≤Dn,∀n∈, \ s.t.\,x_nt_n^loc+ (1-x_n )t_n^tran≤ D_n,∀ n , (7b) fn,min≤fn,mem≤fn,max,∀n∈, \ \ \,f_n, ≤ f_n,mem≤ f_n, ,∀ n , (7c) Fn,min≤fn,com≤Fn,max,∀n∈, \ \ \,F_n, ≤ f_n,com≤ F_n, ,∀ n , (7d) pn,min≤pn≤pn,max,∀n∈, \ \ \,p_n, ≤ p_n≤ p_n, ,∀ n , (7e) xn∈0,1,∀n∈, \ \ \,x_n∈\0,1\,∀ n , (7f) ∑n∈xnBn≤B, \ \ \, _n x_nB_n≤ B, (7g) Bn≥0,∀n∈, \ \ \,B_n≥ 0,∀ n , (7h) where (7b) is the deadline constraint, (7c) is the memory frequency constraint, (7d) is the computing frequency constraint, (7e) is the transmission power constraint, (7f) is the binary offloading decision constraint, (7g) and (7h) denote the bandwidth constraint, respectively. Since the variables xnx_n are binary and coupled in the objective function, Problem P1 is a mixed-integer nonlinear programming (MINLP) problem. To simplify Problem P1, we first focus on the single-user scenario. An optimal solution is solved under specific conditions, while a near-optimal closed-form solution is derived for more general cases. Moreover, for edge inference, we derive an optimal closed-form solution for the transmission power with the given bandwidth. Finally, we propose a heuristic algorithm to solve Problem P1, which greedily searches possible devices to execute edge inference. The detailed analysis and solution are presented as follows. First, consider a special single-user case of Problem P1, and the corresponding problem is formulated as P2:minxn,pn,Bn,fn,mem,fn,comxnenloc+(1−xn)entran 2: _x_n,p_n,B_n,f_n,mem,f_n,comx_ne_n^loc+ (1-x_n )e_n^tran (8a) s.t.xntnloc+(1−xn)tntran≤Dn,∀n∈, \ s.t.\,x_nt_n^loc+ (1-x_n )t_n^tran≤ D_n,∀ n , (8b) fn,min≤fn,mem≤fn,max,∀n∈, \ \ \,f_n, ≤ f_n,mem≤ f_n, ,∀ n , (8c) Fn,min≤fn,com≤Fn,max,∀n∈, \ \ \,F_n, ≤ f_n,com≤ F_n, ,∀ n , (8d) pn,min≤pn≤pn,max,∀n∈, \ \ \,p_n, ≤ p_n≤ p_n, ,∀ n , (8e) xn∈0,1,∀n∈, \ \ \,x_n∈\0,1\,∀ n , (8f) 0≤Bn≤B,∀n∈. \ \ \,0≤ B_n≤ B,∀ n . (8g) We first consider the case where mobile device n performs local inference (i.e., xn=1x_n=1). Then, we can obtain the following optimization problem: P3:minfn,mem,fn,comenloc 3: _f_n,mem,f_n,come_n^loc (9a) s.t.anfn,mem−bn+cnfn,com−dn≤Dn,∀n∈, \ s.t.\,a_nf_n,mem^-b_n+c_nf_n,com^-d_n≤ D_n,∀ n , (9b) fn,min≤fn,mem≤fn,max,∀n∈, \ \ \,f_n, ≤ f_n,mem≤ f_n, ,∀ n , (9c) Fn,min≤fn,com≤Fn,max,∀n∈. \ \ \,F_n, ≤ f_n,com≤ F_n, ,∀ n . (9d) Notice that the constraints (9b), (9c) and (9d) are convex, but the objective function is not always convex. However, by simple computation, we can prove that the objective function of Problem P3 is convex if 0≤bn≤20≤ b_n≤ 2 and 0≤dn≤20≤ d_n≤ 2. In this specific case, Problem P3 is a standard convex optimization problem, and the optimal solutions can be solved by optimization tools (e.g., CVX [20]). Moreover, we try to analyze Problem P3 and give a closed-form solution under a general case. Since the inference time tnt_n should be no larger than the deadline DnD_n, the upper bound of the objective function (9a) is given by replacing tnt_n with DnD_n. The corresponding optimization problem is formulated as P4:minfn,mem,fn,com(κn,memfn,mem3+κn,comfn,com3+σn)Dn 4: _f_n,mem,f_n,com ( _n,memf_n,mem^3+ _n,comf_n,com^3+ _n )D_n (10a) s.t.(9b),(9c),(9d), \ s.t.\,( eq:Problem3b),( eq:Problem3c),( eq:Problem3d), (10b) which is a standard convex optimization problem. Based on the Karush-Kuhn-Tucker (KKT) conditions [21], we can derive the optimal closed-form solutions to Problem P4. The details are given in the following lemma. Lemma 1: The optimal memory frequency and computing frequency to Problem P4 are given by fn,mem∗=max(fn,min,f~n,mem), f_n,mem = (f_n, , f_n,mem), (11) fn,com∗=max(Fn,min,f~n,com), f_n,com = (F_n, , f_n,com), (12) where f~n,mem f_n,mem, f~n,com f_n,com and the Lagrange multiplier λ~n λ_n should satisfy f~n,mem=(λ~nanbn3Dnκn,mem)1bn+3, f_n,mem= ( λ_na_nb_n3D_n _n,mem ) 1b_n+3, (13) f~n,com=(λ~ncndn3Dnκn,com)1dn+3, f_n,com= ( λ_nc_nd_n3D_n _n,com ) 1d_n+3, (14) anf~n,mem−bn+cnf~n,com−dn=Dn. a_n f_n,mem^-b_n+c_n f_n,com^-d_n=D_n. (15) Proof. Denote λn≥0 _n≥ 0 as the Lagrange multiplier, and the Lagrange function can be written as L(fn,mem,fn,com,λn)= L(f_n,mem,f_n,com, _n)= (κn,memfn,mem3+κn,comfn,com3+σn)Dn+ ( _n,memf_n,mem^3+ _n,comf_n,com^3+ _n )D_n+ λn(anfn,mem−bn+cnfn,com−dn−Dn). _n (a_nf_n,mem^-b_n+c_nf_n,com^-d_n-D_n ). (16) Based on the KKT conditions, we have ∂L∂fn,mem|f~n,mem=0, ∂ L∂ f_n,mem |_ f_n,mem=0, (17) ∂L∂fn,com|f~n,com=0, ∂ L∂ f_n,com |_ f_n,com=0, (18) λ~n(anf~n,mem−bn+cnf~n,com−dn−Dn)=0. λ_n (a_n f_n,mem^-b_n+c_n f_n,com^-d_n-D_n )=0. (19) By simple computation, we can get (13) and (14). Since the frequency is positive, the Lagrange multiplier should satisfy λ~n>0 λ_n>0. Combining with (19), we can get (15). Note that the left side of (15) is monotonically decreasing with the increasing of λ~n λ_n, there exists a unique solution for equation (15). In addition, f~n,mem f_n,mem and f~n,com f_n,com should be in the feasible range, which means that fn,mem∗f_n,mem ( fn,com∗f_n,com ) should take the larger one between fn,minf_n, (Fn,minF_n, ) and f~n,mem f_n,mem (f~n,com f_n,com). Therefore, the optimal memory frequency and computing frequency are given by (11) and (12), respectively. ∎ Subsequently, we focus on the optimization of transmission power with the given bandwidth when xn=0x_n=0, and the optimization problem is expressed as P5: minpnpnsnrn _p_n p_ns_nr_n (20a) s.t.snrn≤Dn,∀n∈, \ s.t.\, s_nr_n≤ D_n,∀ n , (20b) pn,min≤pn≤pn,max,∀n∈. \ \ \,p_n, ≤ p_n≤ p_n, ,∀ n . (20c) Although the constraint (20b) is convex, Problem P5 is not a convex optimization problem due to the non-convexity of (20a). However, the objective function of Problem P5 is a monotonically increasing function of transmission power pnp_n. Based on this observation, we can give an optimal closed-form solution, and the details are shown as follows. Lemma 2: The optimal transmission power for Problem P5 is given by pn∗=max(pn,min,p~n),p_n = (p_n, , p_n), (21) where p~n=(2snbnDn−1)σ2hn. p_n= (2 s_nb_nD_n-1 )σ^2h_n. (22) Proof. Denote f(pn)=pnsnrnf(p_n)= p_ns_nr_n as the objective function of Problem P5, where rn=Bnlog2(1+pnhnσ2)r_n=B_n _2 (1+ p_nh_nσ^2 ). Then, we have ∂f∂pn=snlog2Bnlog(1+pnhnσ2)−pnhnpnhn+σ2log2(1+pnhnσ2). ∂ f∂ p_n= s_n 2B_n (1+ p_nh_nσ^2 )- p_nh_np_nh_n+σ^2 ^2 (1+ p_nh_nσ^2 ). (23) Since log(11+x)≤11+x−1 ( 11+x)≤ 11+x-1, we get log(1+x)≥x1+x (1+x )≥ x1+x. Let x=pnhnσ2x= p_nh_nσ^2, we can prove that ∂f∂pn≥0 ∂ f∂ p_n≥ 0, which shows that (20a) monotonically increases with pnp_n. The optimal value of Problem P5 is achieved at the minimum pnp_n in the feasible range. The deadline constraint (20b) requires that the transmission power should be equal or greater than p~n p_n. In addition, pn∗p_n should satisfy the constraint (20c), taking the larger value between p~n p_n and pn,minp_n, . Therefore, pn∗p_n is given by (21). ∎ Based on the above analysis, we propose a heuristic algorithm to effectively solve Problem P1. In each iteration, the total bandwidth is equally allocated among all offloading devices, and the optimal transmission power is obtained by solving Problem P5. If the maximum transmission power among these devices satisfies the available power budget pn,maxp_n, , Algorithm 1 terminates and the corresponding resource allocation is determined. Otherwise, the device requiring the highest transmission power is assigned to perform local inference, and the algorithm is repeated for the remaining devices. The detailed procedure is summarized in Algorithm 1. Since each iteration involves solving convex optimization problems and the maximum number of iterations is N, our proposed algorithm can solve Problem P1 with polynomial computational complexity. Algorithm 1 Heuristic Algorithm for Solving Problem P1 0: Total bandwidth B, number of devices N, deadline DnD_n, channel gain hjj=1N\h_j\_j=1^N, power budget pj,maxj=1N\p_j, \_j=1^N; 0: Offloading policy xnx_n, transmission power pnp_n, bandwidth BnB_n, memory frequency fn,memf_n,mem, computing frequency fn,comf_n,com; 1: Initialize ←1,2,…,NU←\1,2,…,N\; 2: while true do 3: Nact←||N_act←|U|; 4: Equally allocate bandwidth Bn=B/NactB_n=B/N_act, ∀n∈∀ n ; 5: Compute the optimal pn∗p_n for each device based on Lemma 2; 6: Find the maximum pth=maxj∈pjp_th= _j p_j and corresponding device j⋆j ; 7: if pth≤pn,maxp_th≤ p_n, then 8: break // All devices satisfy the power budget 9: else 10: Remove device j⋆j from U; // Eliminate the infeasible device 11: Solve the near-optimal fn,mem∗f_n,mem and fn,com∗f_n,com based on Lemma 1; // Execute local inference 12: end if 13: if =∅U= then 14: break // No feasible offloading solution 15: end if 16: end while 17: Return U, pn∗p_n , BnB_n, fn,mem∗f_n,mem , fn,com∗f_n,com . TABLE I: Parameters of DNNs and Devices. DNN Model Mobile Device fn,memf_n,mem (GHz) fn,comf_n,com (GHz) ana_n bnb_n cnc_n dnd_n κn,mem _n,mem κn,com _n,com σn _n ResNet152 Jetson TX1 [0.102, 1.6] [0.1536, 0.9984] 0.039 1.552 0.211 0.358 0.152 1.115 2.097 VGG19 0.110 1.296 0.099 0.629 IV Simulation Results We assume that there are N=12N=12 mobile devices in a square area of 400m×400m400~m× 400~m, and the total bandwidth is B=20B=20 MHz. The 3GPP channel model hn=38+30log10lnh_n=38+30 _10l_n [22] is adopted here, where lnl_n (in meters) and hnh_n (in dB) are the distance and path loss, respectively. The transmission power range is [0.1,1][0.1,1] W. ResNet152 and VGG19 are deployed on Jetson TX1 to execute DNN inference on CIFAR-100 dataset [23]. The detailed parameters are shown in Table I. Fig. 3 presents the total energy consumption of mobile devices for optimal and near-optimal solutions under various deadlines for local inference on Jetson TX1 with ResNet152 and VGG19. The upper bound is given by executing local inference with the maximum memory frequency and computing frequency. When the deadline is relatively small, the near-optimal solution can achieve the optimal performance. In this case, the optimal energy consumption is achieved when the inference time is equal to the deadline. However, as the deadline further increases, the performance gap increases. For ResNet152 with Dn=0.32D_n=0.32 s and VGG19 with Dn=0.22D_n=0.22 s, the near-optimal solution leads to 2.5%2.5\% and 2.3%2.3\% errors, respectively. Therefore, the performance of near-optimal solution is guaranteed when the deadline is relatively small, where we can obtain the optimal performance with lower computational complexity. In addition, the energy consumption of near-optimal solution does not exceed the upper bound under various deadlines, which further shows the effectiveness of our proposed near-optimal solution. Fig. 3: Comparison for the local energy consumption of optimal and near-optimal solutions under different deadlines. In Fig. 4, we evaluate the corresponding memory frequency and computing frequency for optimal and near-optimal solutions under different deadlines for local inference. It can be observed that the near-optimal memory frequency and computing frequency decrease as the deadline increases, since expressions (13) and (14) monotonically decrease with the deadline DnD_n. However, the optimal memory frequency and computing frequency do not decrease but remain constant when the deadline exceeds a threshold, which is consistent with the results shown in Fig. 3. In this range, the reduction in power consumption caused by the decrease of frequencies can not compensate for the increase of inference time. Specifically, for ResNet152 with Dn=0.32D_n=0.32 s and VGG19 with Dn=0.22D_n=0.22 s, the gap between optimal and near-optimal memory frequency is 15.4%15.4\% and 13.7%13.7\%, while the computing frequency gap is 20.3%20.3\% and 16.0%16.0\%, respectively. Fig. 4: Comparison for the memory frequency and computing frequency of optimal and near-optimal solutions under different deadlines. Finally, we evaluate our proposed Algorithm 1 under different bandwidth resources and device numbers. For comparison, we consider the following three benchmarks. • Random: Instead of removing the device with the maximum transmission power, randomly remove one device from the set U. • Only Compute [24]: Instead of jointly optimizing the two frequencies, only the computing frequency is optimized in Algorithm 1. • No DVFS: Instead of jointly optimizing the two frequencies, the memory frequency and computing frequency remain at the maximum values in Algorithm 1. In Fig. 5, we present the energy consumption for different solving policies. As the total bandwidth increases, more tasks are offloaded from devices to the edge server, which leads to further improvement in energy consumption for all policies. Compared with the Only Compute and No DVFS policies, our proposed Algorithm 1 can achieve average energy savings of 3.6%3.6\% and 10.4%10.4\%, respectively, demonstrating the advantage of jointly optimizing the memory and computing frequency. Furthermore, the proposed algorithm significantly outperforms the Random policy, since it fails to effectively exploit the channel characteristics. It is also observed that when the number of devices is small or the available bandwidth is sufficiently large, the performance gap among different policies becomes less pronounced. Fig. 5: Energy consumption under various bandwidth resources and devices. V Conclusion In this paper, we have studied the joint optimization of memory frequency, computing frequency, bandwidth and transmission power for energy-efficient DNN inference. By integrating analytical modeling for DNN inference time and convex optimization methods, we derive the optimal solution for the local inference problem under specific cases, and a near-optimal closed-form solution is derived for general cases. With the given bandwidth, an optimal closed-form solution for transmission power is derived for the edge inference problem. Finally, a heuristic algorithm is proposed to solve the problem with low computational complexity. Simulation results further verify the characteristics of analytical solutions, where the proposed near-optimal solution can achieve optimal performance in specific cases, and the performance gap with the optimal solution is no more than 2.5%2.5\%. Compared with other methods, our proposed algorithm can substantially reduce the energy consumption of mobile devices. VI Acknowledgement This work is supported in part by the Open Fund of State Key Laboratory of Intelligent Green Vehicle and Mobility, Tsinghua University, and in part by the State Key Laboratory of Internet of Things for Smart City (University of Macau) Open Research Project under Grant SKL-IoTSC(UM)/ORP04/2026, and in part by the Project of Tsinghua University-Toyota Joint Research Center for AI Technology of Automated Vehicle under Grant TTAD-2025-08. References [1] W. Liu, Z. Wang, X. Liu, N. Zeng, Y. Liu, and F. E. Alsaadi, “A survey of deep neural network architectures and their applications,” Neurocomputing, vol. 234, p. 11–26, 2017. [2] Z. Cao et al., “Edge-cloud collaborated object detection via bandwidth adaptive difficult-case discriminator,” IEEE Trans. Mobile Comput., vol. 24, no. 2, p. 1181-1196, Feb. 2025. [3] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, “A survey on mobile edge computing: The communication perspective,” IEEE Commun. Surveys Tuts., vol. 19, no. 4, p. 2322-2358, 4th Quart., 2017. [4] Z. Zhou, X. Chen, E. Li, L. Zeng, K. Luo, and J. Zhang, “Edge intelligence: Paving the last mile of artificial intelligence with edge computing,” Proc. IEEE, vol. 107, no. 8, p. 1738-1762, Aug. 2019. [5] Y. Han, Z. Nan, S. Zhou, and Z. Niu, “DVFS-aware DNN inference on GPUs: Latency modeling and performance analysis,” in Proc. IEEE Int. Conf. Commun. (ICC), Jun. 2025, p. 1274-1279. [6] J. M. Rabaey, A. P. Chandrakasan, and B. Nikolić, Digital Integrated Circuits: A Design Perspective, vol. 7. Upper Saddle River, NJ, USA: Pearson, 2003. [7] Z. Nan, Y. Han, S. Zhou, and Z. Niu, “Robust DNN partitioning and resource allocation under uncertain inference time,” IEEE Trans. Mobile Comput., vol. 25, no. 3, p. 3680-3696, Mar. 2026. [8] W. Shi, S. Zhou, Z. Niu, M. Jiang, and L. Geng, “Multiuser co-inference with batch processing capable edge server,” IEEE Trans. Wireless Commun., vol. 22, no. 1, p. 286-300, Jan. 2023. [9] S. M. Nabavinejad, S. Reda, and M. Ebrahimi, “Coordinated batching and DVFS for DNN inference on GPU accelerators,” IEEE Trans. Parallel Distrib. Syst., vol. 33, no. 10, p. 2496-2508, Oct. 2022. [10] Y. Han, Z. Nan, S. Zhou, and Z. Niu, “Joint memory frequency and computing frequency scaling for energy-efficient DNN inference,” in Proc. IEEE Global Commun. Conf. (GLOBECOM), Dec. 2025, p. 6105-6110. [11] Z. Zhang, Y. Zhao, H. Li, C. Lin, and J. Liu, “DVFO: Learning-based DVFS for energy-efficient edge-cloud collaborative inference,” IEEE Trans. Mobile Comput., vol. 23, no. 10, p. 9042-9059, Oct. 2024. [12] Q. Wang and X. Chu, “GPGPU performance estimation with core and memory frequency scaling,” IEEE Trans. Parallel Distrib. Syst., vol. 31, no. 12, p. 2865-2881, Dec. 2020. [13] B. W. Denkinger et al., “Impact of memory voltage scaling on accuracy and resilience of deep learning based edge devices,” IEEE Des. Test, vol. 37, no. 2, p. 84-92, Apr. 2020. [14] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2016, p. 770-778. [15] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” in Proc. 3rd Int. Conf. Learn. Represent. (ICLR), May 2015, p. 1-14. [16] T. X. Tran and D. Pompili, “Joint task offloading and resource allocation for multi-server mobile-edge computing networks,” IEEE Trans. Veh. Technol., vol. 68, no. 1, p. 856-868, Jan. 2019. [17] Z. Nan, Y. Han, J. Yan, S. Zhou, and Z. Niu, “Robust task offloading and resource allocation under imperfect computing capacity information in edge intelligence systems,” IEEE Trans. Mobile Comput., vol. 24, no. 7, p. 6154-6167, Jul. 2025. [18] J. Haj-Yahya, A. Mendelson, Y. B. Asher, and A. Chattopadhyay, Energy Efficient High Performance Processors: Recent Approaches for Designing Green High Performance Computing. New York, U.S: Springer, 2018. [19] J. Guerreiro, A. Ilic, N. Roma, and P. Tomás, “Modeling and decoupling the GPU power consumption for cross-domain DVFS,” IEEE Trans. Parallel Distrib. Syst., vol. 30, no. 11, p. 2494-2506, Nov. 2019. [20] M. Grant and S. Boyd, “CVX: MATLAB software for disciplined convex programming,” Sep. 2013, [Online]. Available: http://cvxr.com/cvx. [21] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.: Cambridge Univ. Press, 2004, p. 243–247. [22] 3GPP. (Apr. 2022). TR 36.931: Radio Frequency (RF) Requirements for LTE Pico Node B. Version 17.0.0. Accessed: Oct. 2025. [Online]. Available: https://portal.3gpp.org/desktopmodules/Specifications/SpecificationDetails.aspx?specificationId=2589 [23] A. Krizhevsky, “Learning multiple layers of features from tiny images,” M.S. thesis, Univ. Toronto, Toronto, ON, CA, 2009. [24] Z. Nan, S. Zhou, Y. Jia, and Z. Niu, “Joint task offloading and resource allocation for vehicular edge computing with result feedback delay,” IEEE Trans. Wireless Commun., vol. 22, no. 10, p. 6547-6561, Oct. 2023.