Electronics Letters最新文献

Multi-head latent attention (MLA), introduced in DeepSeek-V2, improves the efficiency of large language models by projecting query, key and value tensors into a compact latent space. This architectural change reduces the KV-cache size and significantly lowers memory bandwidth demands, particularly in the autoregressive decode phase. This letter presents the first hardware-centric analysis of MLA, comparing it to conventional multi-head attention (MHA) and evaluating its implications for accelerator performance. We identify two alternative execution schemes of MLA-reusing, respectively recomputing latent projection matrices—which offer distinct trade-offs between compute and memory access. Using the Stream design space exploration framework, we model their throughput and energy cost across a range of hardware platforms and find that MLA can shift attention workloads toward the compute-bound regime. Our results show that MLA not only reduces bandwidth usage but also enables adaptable execution strategies aligned with hardware constraints. Compared to MHA, it provides more stable and efficient performance, particularly on bandwidth-limited hardware platforms. These findings emphasize MLA's relevance as a co-design opportunity for future AI accelerators.

RETRACTION: J. Huang, and S. Lale, “A Novel Nano-Scale Architecture of Vedic Multiplier Using Majority Logic in Quantum-Dot Cellular Automata Technology,” Electronics Letters 58, no 17 (2022): 660–662, https://doi.org/10.1049/ell2.12552.

The above article, published online on 13 June 2022 in Wiley Online Library (wileyonlinelibrary.com), has been retracted by agreement between the journal Editor-in-Chief, Paolo S. Crovetti; The Institution of Engineering and Technology; and John Wiley & Sons Ltd.

The retraction has been agreed due to concerns raised by the first author that they were not involved with the article in any capacity. Upon investigation, we also identified some unusual changes in the submitting author's email and ORCID IDs. We contacted the submitting author to clarify the concerns, but received no response. As we cannot verify the authorship of the publication and have serious concerns about the accountability of the research, we have taken the decision to retract the article. The authors have been informed of the decision to retract.

This letter proposes a convolutional neural network–bidirectional long short-term memory (CNN-BiLSTM) architecture for continuous phase modulation (CPM) signal detection by optimising the extraction of time-frequency features and temporal dependencies with reduced complexity. It significantly outperforms the existing maximum likelihood sequence detection (MLSD) and CNN with fully connected layer (CNN-FC) detectors in higher-order modulation and multipath scenarios, achieving a 97.99% parameter reduction compared to CNN-FC. Numerical results confirm its exceptional balance of detection performance and computational efficiency, making it ideal for complex channels and resource-constrained systems.

This work proposes a 6T1R non-volatile SRAM (nvSRAM) cell based on resistive memory (RRAM) with a small area overhead and low store power compared to previous designs. It features (1) reusing the transistors in the SRAM cell for accessing the RRAM cell, (2) a voltage-division (VD)-based restore process with reduced DC current and (3) a trimmable multi-cycle (TMC) store process to reduce data backup and recovery errors. We fabricated a 1 kb VD-6T1R nvSRAM test array with back-end-of-line integrated metal oxide RRAM cells in a 180 nm CMOS process. The reuse of transistors allows the VD-6T1R cell structure to occupy only 1.14× the area of a standard 6T SRAM cell. The store and restore operations were experimentally verified at the array level. The restore error rates of the fabricated test array can be effectively suppressed using TMC store cycles. The restore errors in the fabricated 1 kb cell array can be eliminated after five cycles.

Digital-domain self-interference cancellation (SIC) technology effectively mitigates self-interference by reconstructing the SI signal. Due to the high computational complexity of the traditional recursive least squares (RLS) algorithm in SIC, its real-time engineering application capability is limited. To overcome this limitation, this paper proposes the application of the DCD-RLS algorithm to SIC. The proposed method transforms the standard normal equations of the RLS algorithm into auxiliary normal equations and incorporates a bisection search strategy, thereby significantly reducing the computational complexity while maintaining optimal convergence performance. Simulation results demonstrate that the proposed algorithm not only achieves a significant reduction in computational complexity but also retains SIC performance, thus enhancing the implementation value of in-band full-duplex (IBFD) communication systems.We propose an RLS-SIC algorithm based on dichotomous coordinate descent (DCD), that converts the RLS normal equations into auxiliary equations and uses a bisection search to greatly reduce computational complexity while preserving optimal convergence. Simulations show, the method maintains self-interference cancellation performance with much lower complexity, improving the real-time feasibility of in-band full-duplex communication systems.

Based on voltage gated spin orbit torque (VG-SOT) MTJ and conventional CMOS circuits, a hybrid ring oscillator (RO) with high power-efficiency and low jitter is proposed in this brief. When the oscillation frequency is 354.1 MHz, the power consumption is only 1.44 mW, and rms jitter is 735.72 fs, results in 1.8$�\times$ and 9$�\times$ of improvements, compared with the 2-terminal MTJ-based hybrid RO.

High-accuracy direction of arrival (DOA) estimation remains challenging for unmanned aerial vehicle (UAV) array platforms due to stringent payload constraints. In this paper, we propose a motion-based synthetic aperture strategy in which a linear array is controlled to move uniformly and sample multiple single snapshots within the temporal coherence period (TCP), thereby enabling two-dimensional (2D) DOA estimation under such restrictive conditions. Furthermore, an Improved Grey Wolf Optimisation (IGWO)-based calibration method is introduced to correct position errors induced by array motion, ensuring accurate angle estimation.

A digital calibration method is presented to reduce SAR ADC offset mismatch. By reusing charge injection (CI) cells, the offset in the sense amplifier (SA) is reduced to less than 1 LSB. In addition, four auxiliary CI cells are incorporated in the digital-to-analogue converter to further reduce the SA offset below 1/2 LSB. Measurement results from a 180 nm CMOS test chip validate the effectiveness of the proposed method. The offset is reduced to 0.5 mV measured at a common-mode voltage of 1.5 V with a 1.8 V power supply.

In pulsed Doppler (PD) radar, active jamming recognition faces two key challenges: complex jamming in scenarios with moving targets and the inefficiency of full-precision (FP) convolutional neural networks (CNNs) for deployment. This letter proposes a binary pyramid self-supervised reconstruction network (BPSR-Net), a lightweight recognition algorithm based on binary neural networks (BNNs). BPSR-Net employs a piecewise quadratic gradient approximation (PQGA) to improve training stability, a multi-scale binary pyramid (BiPy) structure for efficient feature extraction, and a binary self-supervised reconstruction (BiSR) module to enhance global feature learning. Experiments on 12 single and compound jamming types show that BPSR-Net achieves 98.93% top-1 accuracy, 99.90% top-5 accuracy, with only 15.64 Mbit of memory usage and 43.76 × 10⁶ FLOPs. The proposed model balances accuracy, memory, and computational cost, providing a practical solution for low-power radar anti-jamming systems.

This letter presents a 128-element millimetre-wave integrated sensing and communication (ISAC) system. The proposed system integrates the radio frequency (RF) front-end, RF-to-intermediate frequency (IF) converters, IF, and local oscillator (LO) modules into a compact chassis. Multi-stage gain enables a total transmit gain of up to 62 dB and a receive gain of up to 73 dB. A reconfigurable chassis design allows the spacing between the transmit and receive arrays to be flexibly adjusted up to 16 cm, effectively enhancing isolation to 63 dB. When the transmit and receive arrays are co-located, the system supports a wide beam scanning range of ±60° in the azimuth plane and ±50° in the elevation plane and achieves a typical effective isotropic radiated power (EIRP) of 66 dBm at the P1 dB point. Experimental results validate reliable transmission of a 450 MHz 64-QAM signal with an error vector magnitude (EVM) of 4.98%, a maximum sensing range of approximately 350 m, and a range resolution of 34 cm.