Through-silicon vias (TSVs) are essential for 3D integrated circuits (ICs) and advanced chiplet packaging. The semiconductor industry is transitioning toward 3D ICs, chiplets, and system-in-package (SiP) solutions due to the slowdown of Moore’s Law and limitations in conventional silicon scaling. In this paper, we propose an optimized TSV architecture for high-frequency transmission to enhance its suitability for 6G communication chips, and develop a comprehensive equivalent circuit model for fault-free and faulty TSVs. This model accounts for open-circuit and short-circuit fault conditions while considering the effects of higher frequencies, substrate type, doping concentration, and adjacent layers. At the physical level, the TSVs are simulated using the Ansys High-Frequency Structure Simulator (HFSS), and the equivalent circuits are designed using the Cadence Virtuoso tool. An experimental evaluation is also conducted to validate the physical design. We position the TSVs in a pattern of ground-signal-ground (G-S-G) to reduce the effective inductance of the signal TSV, thereby minimizing inductive reactance at ultra-high frequencies. Consequently, the reflection coefficient remains below -10 dB across the frequency range of 0.1 to 146.3 GHz. Furthermore, we compare simulation outcomes from HFSS and Cadence for both fault-free and faulty TSVs under varying operating conditions. Additionally, the parasitic circuit components are characterized through extensive theoretical derivations for in-depth circuit verification. Collectively, the rigorous analysis, experimental validation, and thorough investigation of the proposed design and its equivalent circuit demonstrate their potential for use in creating datasets for a fault prediction machine learning model.

For over half a century, silicon-based Complementary Metal Oxide Semiconductors (CMOS) have been the dominant technology used in the manufacture of nearly all integrated circuits. To fulfill Moore’s Law prediction, CMOS device dimensions were meticulously and carefully reduced to the single-digit nanometer regime. This gradual reduction has led to remarkable exponential performance increases over several decades, ushering in an unparalleled era of computation in human history. However, Short-Channel Effects (SCEs) present many impediments to further improvements in CMOS devices. SCEs occur when the channel length has been scaled down to the same order of magnitude as the depletion-layer widths of the source and drain junctions. To overcome the numerous SCEs caused by the miniaturization of CMOS devices, Carbon Nanotube Field Effect Transistors (CNFETs) aim to serve as their superior successors. CNFETs exhibit exceptional electrical properties, far surpassing those of CMOS, primarily due to their ballistic transport properties and excellent electrostatic scaling. To the best of our knowledge, this paper is the first to investigate chirality variation in CNFETs for spiking neurons. More specifically, chirality variation in CNFETs is used to determine two optimizations: (1) highest-frequency and (2) lowest-energy spiking neurons, using the Penta-Transistor Integrate & Fire (PTIF) architecture to demonstrate these dual mandates. These optimizations separately provide a 6.89x increase in spiking frequency or an 87.43% energy saving.

Stochastic computing (SC) leverages random bitstreams to perform arithmetic operations, offering ultra-low-cost, fault-tolerant, and highly parallelizable computations. The quality of these bit-streams is crucial for the accuracy and reliability of SC. This paper introduces TRUE-BSG, a novel true random bit-stream generator designed for fast and energy-efficient SC. Unlike state-of-the-art (SoTA) pseudo-random and quasi-random bit-stream generators, TRUE-BSG utilizes a high-quality true random number generator (TRNG), capable of producing random bits at a rate of 1 Gigabit per second. Our TRNG ensures high entropy and minimal correlation. TRUE-BSG shows comparable accuracy to software-based generators and better energy efficiency than SoTA bit-stream generators, making it an ideal solution for resource-constrained devices.

Federated Learning (FL) is an emerging paradigm in machine learning without exposing clients' raw data. In practical scenarios with numerous clients, encouraging fair and efficient client participation in federated learning is of utmost importance, which is also challenging given the heterogeneity in data distribution and device properties. Existing works have proposed different client-selection methods that consider fairness; however, they fail to select clients with high utilities while simultaneously achieving fair accuracy levels. In this paper, we propose a fair client-selection approach that unlocks threefold fairness in federated learning. In addition to having a fair client-selection strategy, we enforce an equitable number of rounds for client participation and ensure a fair accuracy distribution over the clients. The experimental results demonstrate that FedFair3, in comparison to the state-of-the-art baselines, achieves 18.15% less accuracy variance on the IID data and 54.78 % on the non-IID data, without decreasing the global accuracy. Furthermore, it shows 24.36% less wall-clock training time on average.

The rapid advancements in nanotechnology and neuromorphic engineering have paved the way for developing high-performance computational models that mimic biological neural networks. This paper presents a novel digital-twin architecture for a spiking neuron, leveraging the exceptional properties of Carbon Nanotube Field Effect Transistors (CNFETs). The proposed architecture aims to emulate the dynamic behavior of biological neurons with high fidelity, providing a robust platform for simulation and analysis. By integrating CNFETs, we achieve significant improvements in lowest-energy usage and highest-spiking frequency, compared to traditional silicon-based technologies. Furthermore, the digital twin not only replicates the electrical characteristics of a spiking neuron but also facilitates advanced functionalities such as adaptive learning, fault tolerance, and self-optimization. Experimental results demonstrate the potential of CNFET-based neurons in enhancing the performance of neuromorphic systems, offering promising applications in artificial intelligence, cognitive computing, and complex system modeling. Integrating digital twin technology with CNFETs allows for precise control and manipulation of spiking behavior, enabling the development of more sophisticated and efficient neural models. This research underscores the transformative impact of combining digital twin technology with cutting-edge nanomaterials, setting a new benchmark for future explorations in the field. The findings highlight the importance of continued research and development in this area, as it holds the promise of revolutionizing the way we design and implement neuromorphic systems, pushing the boundaries of what is possible in artificial intelligence and beyond.

Precise crop yield prediction provides valuable information for agricultural planning and decision-making processes. However, timely predicting crop yields remains challenging as crop growth is sensitive to growing season weather variation and climate change. In this work, we develop a deep learning-based solution, namely Multi-Modal Spatial-Temporal Vision Transformer (MMST-ViT), for predicting crop yields at the county level across the United States, by considering the effects of short-term meteorological variations during the growing season and the long-term climate change on crops. Specifically, our MMST-ViT consists of a Multi-Modal Transformer, a Spatial Transformer, and a Temporal Transformer. The Multi-Modal Transformer leverages both visual remote sensing data and short-term meteorological data for modeling the effect of growing season weather variations on crop growth. The Spatial Transformer learns the high-resolution spatial dependency among counties for accurate agricultural tracking. The Temporal Transformer captures the long-range temporal dependency for learning the impact of long-term climate change on crops. Meanwhile, we also devise a novel multi-modal contrastive learning technique to pre-train our model without extensive human supervision. Hence, our MMST-ViT captures the impacts of both short-term weather variations and long-term climate change on crops by leveraging both satellite images and meteorological data. We have conducted extensive experiments on over 200 counties in the United States, with the experimental results exhibiting that our MMST-ViT outperforms its counterparts under three performance metrics of interest. Our dataset and code are available at https://github.com/fudong03/MMST-ViT.