Artificial intelligence infrastructure is scaling rapidly, yet assessments of its environmental footprint have focused almost exclusively on operational electricity consumption. Here we show that this framing omits a large and growing contribution: the energy embodied in manufacturing the hardware itself. Analyzing datacenter AI accelerators, we find that high-bandwidth memory (HBM) fabrication alone accounts for 55–71% of total manufacturing energy because sustaining AI throughput requires stacking dozens of memory dies per device. Projecting vendor roadmaps through 2035, manufacturing energy per accelerator is estimated to grow six- to nine-fold as HBM scales toward 800 GB–1.5 TB. The architectural advances that reduce operational energy per token, such as precision quantization, attention compression, and sparse mixture-of-expert activation, act on compute and memory-bandwidth utilization rather than on memory capacity, leaving manufacturing energy unaffected by the same innovations that improve operational efficiency. These results demonstrate that hardware sustainability assessments should account explicitly for embodied manufacturing energy alongside operational consumption.

Triboelectric nanogenerators (TENGs) are attracting increasing attention as viable power sources for self-powered systems, due to advantages such as material and form-factor versatility, compatibility with low-frequency mechanical stimuli, and scalable, low-cost fabrication. Unlike conventional harvesters, TENGs exhibit two device-level characteristics that critically shape interface-circuit design: (i) a time-varying internal capacitance that induces a strongly dynamic source impedance, and (ii) a significantly high effective source impedance that yields very high open-circuit voltages at low currents. These characteristics directly affect impedance matching, rectification, voltage conversion, and maximum power point tracking (MPPT) strategies for maximum power transfer. While TENGs and piezoelectric energy harvesters (PEHs) share similar lumped electrical models, the time variance and voltage/current operating regime of TENGs fundamentally limit the portability of PEH-oriented power transfer methods. This paper provides two contributions. First, we introduce a figure-of-merit (FoM) that serves as an energy-extraction coefficient: the fraction of the ideal maximum power of the TENG device (under instantaneous impedance tracking) that appears at the rectifier input, typically the first stage of a power management unit (PMU). The FoM exposes losses arising from mismatch at the device-PMU boundary, thereby helping circuit designers localize dominant loss mechanisms (e.g., impedance mismatch, rectifier topology or suboptimal MPPT policies) and guiding device researchers to prioritize physical parameters (e.g., dielectric thickness, displacement, electrode area) with explicit awareness of interface constraints. Secondly, we conduct a thorough evaluation of the suitability of advanced PEH-derived methods, such as rectifiers (both passive and active configurations), DC-DC conversion, and MPPT, for application to TENGs. We derive theoretical upper bounds on extractable power for representative rectifier families under TENG-specific operating conditions, and we analyze technology-imposed voltage limits and their implications for architecture and control. We also survey recent TENG demonstrations together with their PMU interfaces and interpret reported performance through the proposed FoM. Overall, the analysis highlights that the unique characteristics of TENGs and technology limits for the voltages must be explicitly accounted for while developing interface circuits to realize maximum power extraction. This process can significantly benefit from coordinated device-circuit co-design via system-level metrics such as the proposed FoM.

The rapid growth of Deep Neural Networks (DNNs) has led to the development of application-specific DNN accelerators. Conventional 2D von Neumann architectures suffer from memory bandwidth limitations between the memory and the processing core. 3D DNN accelerators have emerged as a promising solution by leveraging 3D integration to enable near-memory logic or in-memory computation. By shifting computation closer to memory, these accelerators significantly reduce data movement and therefore latency, resulting in more energy-efficient operations. Monolithic 3D (M3D) integration, in particular, enables high-bandwidth systems by utilizing high-density monolithic inter-tier vias (MIVs). This paper provides a critical review of recent advances in 3D DNN accelerators that combine near-memory and compute-in-memory with various 3D technologies, offering a useful discussion and future prospects of the available technologies and architectures that have advanced the performance of DNN accelerators. Particular attention is devoted to accelerators for emerging transformer-based large language model (LLM) networks due to the higher memory demands. Thermal-aware design techniques of 3D DNN accelerators are also discussed as a means to address the fundamental challenge of heat dissipation. A detailed review is finally conducted on package-level constraints, considering signal integrity, power delivery, and thermo-mechanical reliability.

This paper describes a method for maximizing the power delivered to the rectifier in energy harvesters. It is demonstrated that there is an optimal turn-on time for rectifiers to maximize power transfer from the harvester. Next, a maximum power point tracking methodology based on rectifier turn-on time (RTOT-MPPT) is developed for triboelectric energy harvesters. The primary advantage of the proposed approach is the relative independence of the optimal turn-on time on the frequency and peak voltage of the harvester output. Thus, the proposed RTOT-MPPT method reduces the complexity of power tracking and can be efficient for a wider range of harvesters. The method is implemented for a triboelectric harvester and simulated in a 180 n m industrial HV-CMOS process, demonstrating that 34% higher power is delivered to the rectifier in each cycle.

Developing self-powered, durable pressure sensors for Total Knee Replacement (TKR) enhances longevity, ensures consistent performance, and provides critical post-operative information. This study presents a triboelectric nanogenerator (TENG) integrated into an instrumented knee implant for energy harvesting and pressure sensing. Operating in vertical contact mode, it utilizes porous silicone rubber (SR) dielectric to improve electrical stability, and mechanical durability. The nanogenerator divides the tibial tray into two compartments for load imbalance detection. Tests simulating human walking showed the device withstands forces up to 2200N, generating a maximum of 18 mu W at 1Hz under harmonic load and a maximum of 7.5 mu W at 0.8Hz under gait loading with a VIVO joint simulator. The performance of the TENG was stable over 3000 cycles generating a peak-to-peak voltage of 350V. The porous structure enhances charge trapping, energy storage, and system efficiency. The increased power compared to previous work enhances energy harvesting capability and strengthens its potential for self-powered, real-time load monitoring at the knee joint.

This paper describes an analytic method to determine the optimal load capacitance of a full wave rectifier (FWR) in triboelectric energy harvesters. The amount of average power delivered from the harvester to the rectifier grows in each cycle, eventually reaching a peak value. In steady state, the average rectifier power depends on the harvester characteristics and the load capacitance of the rectifier. For a given harvester, if the load capacitance is too small, the rectifier power does not reach the maximum power in steady state. Alternatively, if the load capacitance is too large, the number of cycles to reach maximum power increases, causing additional delay. An analytic method is proposed to estimate the optimum value of this capacitance, which ensures maximum power while minimizing the transient time. The results are validated with the measurement results of a triboelectric harvester.

This paper proposes a co-design method that integrates the central processing unit (CPU) and resistive random-access memory (ReRAM) crossbar to enhance energy efficiency of Octave convolution. In this method, low-frequency operations within the Octave framework are executed by the energy-efficient ReRAM crossbar, while high-frequency computations, which are crucial for accuracy, are processed by the CPU and results are combined within the CPU to finalize the classification task. Compared to vanilla convolution, for ResNet-50 trained on the CIFAR-10 dataset, this approach reduces the number of CPU operations by 37% while improving accuracy by 4.7%. Compared to traditional Octave convolution, the proposed approach reduces the number of CPU operations by 17% while maintaining accuracy within 2%. Approximately 18.5% of the total operations is executed by the energy-efficient ReRAM in the proposed approach.

Although total knee replacements have an insignificant impact on patients’ mobility and quality of life, real-time performance monitoring remains a challenge. Monitoring the load over time can improve surgery outcomes and early detection of mechanical imbalances. Triboelectric nanogenerators (TENGs) present a promising approach as a self-powered sensor for load monitoring in TKR. A TENG was fabricated with dielectric layers consisting of Kapton tape and 3D-printed thermoplastic polyurethane (TPU) matrix incorporating CNT and BTO fillers, separated by an air gap and sandwiched between two copper electrodes. The sensor performance was optimized by varying the concentrations of BTO and CNT to study their effect on the energy-harvesting behavior. The test results demonstrate that the BTO/TPU composite that has 15% BTO achieved the maximum power output of 11.15 μW, corresponding to a power density of 7 mW/m2, under a cyclic compressive load of 2100 N at a load resistance of 1200 MΩ, which was the highest power output among all the tested samples. Under a gait load profile, the same TENG sensor generated a power density of 0.8 mW/m2 at 900 MΩ. By contrast, all tested CNT/TPU-based TENG produced lower output, where the maximum generated apparent power output was around 8 μW corresponding to a power density of 4.8 mW/m2, confirming that using BTO fillers had a more significant impact on TENG performance compared with CNT fillers. Based on our earlier work, this power is sufficient to operate the ADC circuit. Furthermore, we investigated the durability and sensitivity of the 15% BTO/TPU samples, where it was tested under a compressive force of 1000 N for 15,000 cycles, confirming the potential of long-term use inside the TKR. The sensitivity analysis showed values of 37.4 mV/N for axial forces below 800 N and 5.0 mV/N for forces above 800 N. Moreover, dielectric characterization revealed that increasing the BTO concentration improves the dielectric constant while at the same time reducing the dielectric loss, with an optimal 15% BTO concentration exhibiting the most favorable dielectric properties. SEM images for BTO/TPU showed that the 10% and 15% BTO/TPU composites showed better morphological characteristics with lower fabrication defects compared with higher filler concentrations. Our BTO/TPU-based TENG sensor showed robust performance, long-term durability, and efficient energy conversion, supporting its potential for next-generation smart total knee replacements.

We propose a power management strategy that maximizes the power harvested from a triboelectric nanogenerator using a Parallel Synchronous Switched Harvesting on Inductor (P-SSHI)-based rectifier. By analyzing the charge transfer from the rectifier to the DC-DC converter, we observe increase in the extracted power by limiting the rectifier capacitor discharging. We design and implement a power management system based on the proposed switching technique. The technique is suitable for integration into a power management system of smart selfpowered sensors, such as those used in smart knee implants following total knee replacement (TKR) surgery. We demonstrate 37% increase in the harvested power at the TENG interface compared to the conventional switching in P-SSHI-based rectifier.

This study investigates the energy harvesting and sensing capabilities of piezoelectric nanogenerators (PENG) and triboelectric nanogenerators (TENG) for long-term load monitoring in total knee replacement (TKR). Multi-layered polyvinylidene fluoride (PVDF) films and cuboid-patterned silicone rubber embedded with dopamine-coated BaTiO3 particles (SR/BT@PDA) TENG are compared as energy harvesting-based load sensors. Unlike prior studies relying on simplified harmonic loading, this work utilizes physiologically relevant gait cycles covering realistic force ranges to precisely evaluate electrical output, sensitivity, and activity recognition capabilities. Results indicate forward-polarized TENG samples and upward-polarized PVDF layers generate significantly higher outputs, indicating the importance of dipole alignment for enhanced sensor efficiency. The harvesters' outputs show that the SR/BT@PDA TENG achieves a maximum apparent power output of 6 μW at 1.5GΩ, while the PVDF reaches 2.7 μW at 200MΩ under normal walking conditions. The SR/BT@PDA TENG outperforms PVDF in energy harvesting, reaching 140 V in 26 gait cycles for a 10nF capacitor and powering 60 LEDs, while PVDF charges the same capacitor to 33 V in nearly 19 gait cycles, powering 14 LEDs. The TENG's micro-cuboid surface patterning and synergistic effects of embedded piezoelectric material (BaTiO3) enhance its output power density, whereas the multi-layered PVDF demonstrates reliable performance under diverse load conditions. Both sensors effectively detect diverse activities, including walking, jogging, and stair climbing. Overall, PVDF provides precise load monitoring by tracking dynamic force profiles, while TENG outperforms in energy harvesting. This study evaluates the potential of integrating TENG and PENG into TKR as energy-harvesting solutions for joint load monitoring without relying on external power sources.

A power optimization strategy is described for triboelectric energy harvesting systems by optimizing the load capacitor size within a full wave rectifier (FWR). In AC harvesters such as triboelectric nanogenerators (TENGs) with an FWR, the average power delivered to the rectifier increases in each cycle, ultimately reaching a steady state determined by system parameters. Through cycle-level analysis of input voltage, current, and rectifier turn-on time during mechanical motion, an optimal load capacitance is identified that maximizes power delivery while minimizing transient time to reach this maximum power. This approach achieves peak power delivery via capacitor sizing alone, eliminating the need for additional circuitry. Experimental results using a vertical contact-separation triboelectric nanogenerator with internal capacitance varying from 24pF to 96pF demonstrate that the optimal rectifier capacitance of 390pF achieves maximum power delivery (900nW at 1.7Hz and 2.7μW at 5Hz) within the second cycle, while suboptimal capacitances either fail to reach peak power or delay it to the seventh cycle or later. Sensitivity analysis reveals that the method exhibits high robustness, with capacitances within ±30% of the optimal value can still maintain ≥ 90% of peak power, providing flexibility when implementing or selecting the capacitor size.

This study presents the development and characterization of a novel triboelectric nanogenerator (TENG) designed as a self-powered sensor for load monitoring in total knee replacement (TKR) implants. The triboelectric layers comprise a 3D-printed thermoplastic polyurethane (TPU) matrix with carbon nanotube (CNT) nanoparticles and kapton tape, sandwiched between two copper electrodes. To optimize sensor performance, the proposed CNT/TPU TENG sensor is fabricated with varying CNT concentrations and thicknesses, enabling a comprehensive analysis of how material composition and structural parameters influence energy harvesting efficiency. The 1% CNT/TPU composite demonstrates the highest power output among the tested samples. The solid CNT/TPU-based TENG generated the apparent output power of 4.1 µW under a cyclic compressive load of 2100 N, measured across a 1.6 GΩ load resistance and over a nominal contact area of 15.9 cm2, while the foam CNT/TPU film achieved a higher apparent output power of 6.9 µW measured across a 0.9 GΩ load resistance with the same nominal area. The generated power is sufficient to operate a power management and ADC circuit based on our earlier work. The sensors exhibit a stable open-circuit voltage of 320 V for the foam layer and 275 V for the solid one. Sensitivities are 80.50 mV N-1 ( ⩽ 1600  N) and 24.60 mV N-1 (> 1600 N) for foam CNT/TPU film, demonstrating the integrated sensor capability for wide-range force sensing on TKR implants. The foam CNT/TPU-based TENG maintained stable performance over 16 000 load cycles, confirming its potential for long-term use inside the TKR. Additionally, the dielectric constant of the CNT/TPU composite was found to increase with increasing CNT concentration. The proposed CNT/TPU TENG sensor offers a broad working range and robust energy-harvesting efficiency, making it appropriate for self-powered load sensing in biomedical applications.

We propose a peak detection and frequency measurement circuit for integration at interface with a triboelectric nanogenerator (TENG). By measuring the frequency and the peak voltage at the output of TENG, the applied force is predicted. We propose a sample and hold peak detector circuit with a reset signal for continues peak detection. The control logic also generates a trigger signal for analog to digital conversion of the peak voltage to the digital domain. At the same time, the frequency is recorded by counting the number of peaks in a defined time period. The circuit simulations in 180 nm CMOS technology demonstrate peak conversion with 8-bit resolution in a 1 V voltage range with 500 nW power consumption. The proposed design is amenable for integration in a self-powered load sensing system in smart knee implant after total knee replacement(TKR) surgery.

Systolic arrays are popular for executing deep neural networks (DNNs) at the edge. Low latency and energy efficiency are key requirements in edge devices such as drones and autonomous vehicles. Monolithic 3D (MONO3D) is an emerging 3D integration technique that offers ultra-high bandwidth among processing and memory elements with a negligible area overhead. Such high bandwidth can help meet the ever-growing latency and energy efficiency demands for DNNs. This paper presents a novel implementation for weight stationary (WS) dataflow in MONO3D systolic arrays, called WS-MONO3D. WS-MONO3D utilizes multiple resistive RAM layers and SRAM with high-density vertical interconnects to multicast inputs and perform high-bandwidth weight pre-loading while maintaining the same order of multiply-and-accumulate operations as in native WS dataflow. Consequently, WS-MONO3D eliminates input and weight forwarding cycles and, thus, provides up to 40% improvement in energy-delay-product (EDP) over the native WS implementation in 2D at iso-configuration. WS-MONO3D also provides 10X improvement in inference per second per watt per footprint due to multiple vertical tiers. Finally, we also show that temperature impacts the energy efficiency benefits in WS-MONO3D.

A self-powered and durable pressure sensor for large-scale pressure detection on the knee implant would be highly advantageous for designing long-lasting and reliable knee implants as well as obtaining information about knee function after the operation. The purpose of this study is to develop a robust energy harvester that can convert wide ranges of pressure to electricity to power a load sensor inside the knee implant. To efficiently convert loads to electricity, we design a cuboid-array-structured tribo-pizoelectric nanogenerator (TPENG) in vertical contact mode inside a knee implant package. The proposed TPENG is fabricated with aluminum and cuboid-patterned silicone rubber layers. Using the cuboid-patterned silicone rubber as a dielectric and aluminum as electrodes improves performance compared with previously reported self-powered sensors. The combination of 10 w t % dopamine-modified BaTiO3 piezoelectric nanoparticles in the silicone rubber enhanced electrical stability and mechanical durability of the silicone rubber. To examine the output, the package-harvester assemblies are loaded into an MTS machine under different periodic loading. Under different cyclic loading, frequencies, and resistance loads, the harvester's output performance is also theoretically studied and experimentally verified. The proposed cuboid-array-structured TPENG integrated into the knee implant package can generate approximately 15 mu W of apparent power under dynamic compressive loading of 2200 N magnitude. In addition, as a result of the TPENG's materials being effectively optimized, it possesses remarkable mechanical durability and signal stability, functioning after more than 30 000 cycles under 2200 N load and producing about 300 V peak to peak. We have also presented a mathematical model and numerical results that closely capture experimental results. We have reported how the TPENG charge density varies with force. This study represents a significant advancement in a better understanding of harvesting mechanical energy for instrumented knee implants to detect a load imbalance or abnormal gait patterns.

Integrated circuits (ICs) have become increasingly susceptible to counterfeiting due to globalization of the semiconductor supply chain. In this paper, this issue is addressed with a novel IC authentication approach by leveraging both on-chip (transistor and interconnect) and package-level process variations. Unique IDs are generated with this random variation by obtaining the supply-side power network impedances of packaged ICs over a certain frequency range. PowerID requires a simpler measurement setup than other offline signature generation techniques that rely on measuring path delays by generating specific input patterns. Unlike existing methods, such as physically unclonable functions (PUFs) that rely on dedicated circuit structures, this approach does not require any additional circuitry. Simulations are performed to model 1,000 IC instances and common metrics such as uniqueness (average inter-Hamming distance of 49.5%) and robustness (average intra-Hamming distance of 1%) are reported. Due to the random nature of process variations, it is highly challenging for an attacker to deliberately produce a counterfeit IC that matches the generated unique signatures.

Systolic arrays are commonly used for running deep neural networks (DNNs) at the edge, where latency and energy efficiency requirements are stringent. Monolithic 3D (Mono3D) is an emerging 3D integration technology that offers ultra-high vertical interconnect density among processing and memory layers. The bandwidth benefits provided by Mono3D can help meet the growing latency and energy efficiency demands for DNNs. This paper presents a novel implementation for weight stationary (WS) dataflow in Mono3D systolic arrays, called WS-Mono3D. WS-Mono3D utilizes multiple resistive RAM layers and SRAM with high-density vertical interconnects to multicast inputs and performs high-bandwidth weight pre-loading while maintaining the same order of multiply-and-accumulate operations as in native WS dataflow. Consequently, WS-Mono3D eliminates input and weight forwarding cycles, and, thus, provides up to a 40% reduction in energy-delay-product (EDP) over the native WS implementation in 2D with iso-configuration. The paper also demonstrates the impact of temperature on energy efficiency benefits in WS-Mono3D.

A present challenge in structural health monitoring consists in the detection, localization, and quantification of small damage (e.g., small cracks) within large structures, such as bridges and buildings. Existing sensing solutions have several limitations, the most important being those related to the extent of spatial coverage by sensors and power supply. In this work, we will present proof-of-concept research for sub-millimeter displacement measurement using novel embeddable passive wireless radio frequency (RF) sensors. The novel sensors estimate relative displacement from phase shifts in the transmitted RF signal. The proposed system represents a novel paradigm in wireless sensing in structural health monitoring, as the wireless sensors are battery-less and will be deployed in a form of densely populated 3D network embedded within large volume of material.

Although total knee replacements (TKR) are generally considered highly successful, patient satisfaction may not always be sustained over time. Monitoring the performance of the replaced knee can lead to improved surgical outcomes. Triboelectric Nano Generator (TENG) can generate electrical energy to power a load sensor, enabling continuous monitoring of the total knee replacement and more specifically monitoring imbalance over time. The objective of this study is to create a TENG capable of handling body forces and generating enough electrical energy to power the sensors. This is achieved through the energy harvester, which generates electric power by utilizing the contact separation resulting from cyclic compressive loads. To enhance the electrical properties of the energy harvester, a keystroke shape was used as a TENG structure. Moreover, the favorable characteristics of biocompatibility of the materials used in this study carry significant importance, particularly in applications where compatibility with biological systems is essential. This makes them particularly well-suited for integration into total knee implants. The experiment was conducted in this study and showed that the keystroke TENG structure could generate around 20 V peak voltage 2 kN axial force was applied to the TENG at 1 Hz. Expanding this to an array design enables measuring pressure distribution across TKR.

Information leakage through temperature based covert channels is a growing threat in modern multicore processors. Thus, accurately detecting the presence of such thermal covert communication channels in real time is crucial for ensuring the security of confidential data. Existing detection techniques fail when covert channels are implemented with low power benchmarks, as shown in this paper. A novel detection technique is proposed by considering the transitions in the CPU workload as the primary metric. The proposed approach can detect covert channels established with low power programs with 100% detection accuracy and less than 2% false positive rate.