IEEE Transactions on Very Large Scale Integration (VLSI) Systems最新文献

{"title":"Xpikeformer: Hybrid Analog-Digital Hardware Acceleration for Spiking Transformers","authors":"Zihang Song;Prabodh Katti;Osvaldo Simeone;Bipin Rajendran","doi":"10.1109/TVLSI.2025.3552534","DOIUrl":"https://doi.org/10.1109/TVLSI.2025.3552534","url":null,"abstract":"The integration of neuromorphic computing and transformers through spiking neural networks (SNNs) offers a promising path to energy-efficient sequence modeling, with the potential to overcome the energy-intensive nature of the artificial neural network (ANN)-based transformers. However, the algorithmic efficiency of SNN-based transformers cannot be fully exploited on GPUs due to architectural incompatibility. This article introduces Xpikeformer, a hybrid analog-digital hardware architecture designed to accelerate SNN-based transformer models. The architecture integrates analog in-memory computing (AIMC) for feedforward and fully connected layers, and a stochastic spiking attention (SSA) engine for efficient attention mechanisms. We detail the design, implementation, and evaluation of Xpikeformer, demonstrating significant improvements in energy consumption and computational efficiency. Through image classification tasks and wireless communication symbol detection tasks, we show that Xpikeformer can achieve inference accuracy comparable to the GPU implementation of ANN-based transformers. Evaluations reveal that Xpikeformer achieves a <inline-formula> <tex-math>$13times $ </tex-math></inline-formula> reduction in energy consumption at approximately the same throughput as the state-of-the-art (SOTA) digital accelerator for ANN-based transformers. In addition, Xpikeformer achieves up to <inline-formula> <tex-math>$1.9times $ </tex-math></inline-formula> energy reduction compared to the optimal digital ASIC projection of SOTA SNN-based transformers.","PeriodicalId":13425,"journal":{"name":"IEEE Transactions on Very Large Scale Integration (VLSI) Systems","volume":"33 6","pages":"1596-1609"},"PeriodicalIF":2.8,"publicationDate":"2025-04-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144117347","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Flex-PE: Flexible and SIMD Multiprecision Processing Element for AI Workloads","authors":"Mukul Lokhande;Gopal Raut;Santosh Kumar Vishvakarma","doi":"10.1109/TVLSI.2025.3553069","DOIUrl":"https://doi.org/10.1109/TVLSI.2025.3553069","url":null,"abstract":"The rapid evolution of artificial intelligence (AI) models, from deep neural networks (DNNs) to transformers/large-language models (LLMs), demands flexible hardware solutions to meet diverse execution needs across edge and cloud platforms. Existing accelerators lack unified support for multiprecision arithmetic and runtime-configurable activation functions (AFs). This work proposes Flex-PE, a single instruction, multiple data (SIMD)-enabled multiprecision processing element that efficiently integrates multiply-and-accumulate operations with configurable AFs using unified hardware, including Sigmoid, Tanh, ReLU, and SoftMax. The proposed design achieves throughput improvements of up to <inline-formula> <tex-math>$16times $ </tex-math></inline-formula> FxP4, <inline-formula> <tex-math>$8times $ </tex-math></inline-formula> FxP8, <inline-formula> <tex-math>$4times $ </tex-math></inline-formula> FxP16, and <inline-formula> <tex-math>$1times $ </tex-math></inline-formula> FxP32, with maximum hardware efficiency for both iterative and pipelined architectures. An area-efficient iterative Flex-PE-based SIMD systolic array reduces DMA reads by up to <inline-formula> <tex-math>$62times $ </tex-math></inline-formula> and <inline-formula> <tex-math>$371times $ </tex-math></inline-formula> for input feature maps and weight filters in VGG-16, achieving 8.42 GOPS/W energy efficiency with minimal accuracy loss (<2%). Flex-PE scales from 4-bit edge inference to FxP8/16/32, supporting edge and cloud high-performance computing (HPC) while providing high-performance adaptable AI hardware with optimal precision, throughput, and energy efficiency.","PeriodicalId":13425,"journal":{"name":"IEEE Transactions on Very Large Scale Integration (VLSI) Systems","volume":"33 6","pages":"1610-1623"},"PeriodicalIF":2.8,"publicationDate":"2025-04-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144117346","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A 28-nm Cascode Current Mirror-Based Inconsistency-Free Charging-and-Discharging SRAM-CIM Macro for High-Efficient Convolutional Neural Networks","authors":"Chunyu Peng;Jiating Guo;Shengyuan Yan;Yiming Wei;Xiaohang Chen;Wenjuan Lu;Chenghu Dai;Zhiting Lin;Xiulong Wu","doi":"10.1109/TVLSI.2025.3552641","DOIUrl":"https://doi.org/10.1109/TVLSI.2025.3552641","url":null,"abstract":"Computing-in-memory (CIM) is an emerging approach to alleviate the von Neumann bottleneck and enhance energy efficiency and throughput. This brief introduces a 16-Kb static random access memory (SRAM) CIM macro for convolutional neural networks (CNNs), featuring a cascode current mirror-based inconsistency-free computing circuits (CICCs). The bias voltage of CICC is provided by a cascode current mirror (CCM) circuit. The proposed architecture improves the consistency and linearity of bitline (BL) charge and discharge rates in the analog current domain, enhancing computational accuracy. Additionally, the charge and discharge on the BLs represent the positive or negative calculation result, eliminating the need for extra encoding and logic circuits to handle sign bits. The SRAM-CIM macro achieves an energy efficiency of 59.1–134.0 TOPS/W and a throughput of 0.41 TOPS in a 28-nm CMOS technology, and the estimated inference accuracy on MNIST and CIFAR-10 datasets is 96.5% and 91.4%, respectively, with 5-bit input precision and 1-bit weight precision.","PeriodicalId":13425,"journal":{"name":"IEEE Transactions on Very Large Scale Integration (VLSI) Systems","volume":"33 7","pages":"2044-2048"},"PeriodicalIF":2.8,"publicationDate":"2025-04-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144519298","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Cost-Optimized Double-Node-Upset-Recovery Latch Designs With Aging Mitigation and Algorithm-Based Verification for Long-Term Robustness Enhancement","authors":"Aibin Yan;Changli Hu;Jing Li;Na Bai;Zhengfeng Huang;Tianming Ni;Girard Patrick;Xiaoqing Wen","doi":"10.1109/TVLSI.2025.3554117","DOIUrl":"https://doi.org/10.1109/TVLSI.2025.3554117","url":null,"abstract":"With the continuous advancement of CMOS technologies, soft errors, such as single-node upset (SNU) and double-node upset (DNU), caused by radiation in nanoscale integrated circuits, are becoming increasingly prominent. Meanwhile, transistor aging mitigation is indispensable for long-term robustness enhancement. First, to reduce the impact of radiation on circuits, we propose a novel DNU-recovery latch with low cost, namely, DURLC, only consisting of four dual-input C-elements (CEs) and four clock-gated input-split inverters for the storage of values. Second, we propose a DNU-recovery latch with moderate cost, namely, DURMC, based on seven CEs and four inverters, for convenience to optimize the latch to alleviate aging. The proposed DNU-recovery latch with mitigated aging is called DURMA. The latch employs a high-speed path to reduce delay without sacrificing performance when mitigating aging issues. Finally, we propose an algorithm-based verification method to validate the DNU recovery of the proposed latches. The simulation results show that, compared with the state-of-the-art robust latches, the proposed latches have the advantages of DNU recovery with moderate and even low cost, and meanwhile, aging is effectively mitigated for the DURMA latch.","PeriodicalId":13425,"journal":{"name":"IEEE Transactions on Very Large Scale Integration (VLSI) Systems","volume":"33 6","pages":"1765-1773"},"PeriodicalIF":2.8,"publicationDate":"2025-04-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144117272","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Hybrid Number Theoretic Transform Architecture for Homomorphic Encryption","authors":"Quang Dang Truong;Phap Duong-Ngoc;Hanho Lee","doi":"10.1109/TVLSI.2025.3552852","DOIUrl":"https://doi.org/10.1109/TVLSI.2025.3552852","url":null,"abstract":"Fully homomorphic encryption (FHE) is an innovative cryptographic technology that has the potential to protect the privacy and confidentiality of data in the untrusted environments, such as public clouds or external parties. However, due to the inclusion of time-consuming polynomial arithmetic, FHE remains a challenge for computationally heavy applications. The number theoretic transform (NTT) is widely used in HE to reduce the complexity of polynomial multiplication. Therefore, implementing NTT in hardware for FHE has been explored in prior studies. However, due to the high hardware resource requirements, especially with a large number of moduli, hardware architecture supporting both NTT and its inverse transform (INTT) is still missing. This brief presents a hardware architecture for <inline-formula> <tex-math>$2^{17}$ </tex-math></inline-formula> NTT and INTT suitable for high-circuit depth CKKS-based HE schemes, satisfying both criteria of high speed and affordability for various FPGA platforms. The implementation results highlight that this design is area-efficient compared to the most related work and hardware-friendly for practical HE-based applications on FPGA devices.","PeriodicalId":13425,"journal":{"name":"IEEE Transactions on Very Large Scale Integration (VLSI) Systems","volume":"33 7","pages":"2039-2043"},"PeriodicalIF":2.8,"publicationDate":"2025-04-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144519347","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A 2-Lane DAC-/ADC-Based 2 × 2 MIMO PAM-4 MMSE-DFE Wireline Transceiver With FEXT Cancellation on RFSoC Platform","authors":"Jaewon Lee;Seoyoung Jang;Yujin Choi;Donggeon Kim;Matthias Braendli;Thomas Morf;Marcel Kossel;Pier-Andrea Francese;Gain Kim","doi":"10.1109/TVLSI.2025.3553400","DOIUrl":"https://doi.org/10.1109/TVLSI.2025.3553400","url":null,"abstract":"This article presents a 2-lane <inline-formula> <tex-math>$2 times 2$ </tex-math></inline-formula> multiple-input, multiple-output (MIMO) 4-level pulse amplitude modulation (PAM-4) minimum mean-squared-error (MMSE)-decision-feedback equalizer (DFE) with far-end crosstalk (FEXT) cancellation for digital-to-analog converter (DAC)-/analog-to-digital converter (ADC)-based high-speed serial links. The receiver (RX) datapath is designed with a 15-tap MIMO feedforward equalizer (FFE) and a one-tap MIMO DFE with the least mean square (LMS), enabling adaptation to channel variation while maintaining the MMSE setting. The RX digital signal processor (DSP) place and route (PnR) in a 28-nm CMOS is estimated to consume 201 mW/lane at a 56-Gb/s/lane data rate while occupying a 0.5-mm²/lane silicon area. We further implement a real-time evaluation platform to verify the functionality of the MIMO PAM-4 MMSE-DFE with rapid bit-error-rate (BER) testing on RFSoC. The measurement result demonstrates that the MIMO MMSE-DFE significantly improves BER performance from 2.75e⁻³ to 1.31e⁻⁷ compared with equalization without FEXT cancellation when communicating over a channel exhibiting 12.4-dB insertion loss (IL) and 13.2-dB IL-to-crosstalk ratio (ICR) at Nyquist.","PeriodicalId":13425,"journal":{"name":"IEEE Transactions on Very Large Scale Integration (VLSI) Systems","volume":"33 6","pages":"1570-1581"},"PeriodicalIF":2.8,"publicationDate":"2025-04-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144117385","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Revisiting Multiple ECC on High-Density NAND Flash memory","authors":"Yunpeng Song;Yina Lv;Wentong Li;Jialin Liu;Liang Shi","doi":"10.1109/TVLSI.2025.3551400","DOIUrl":"https://doi.org/10.1109/TVLSI.2025.3551400","url":null,"abstract":"Three-dimensional <sc>nand</small> flash memory using the advanced multibit-per-cell technique is widely adopted due to its high density. However, it faces the problem of deteriorating read performance and energy consumption due to decreased reliability. Low-density parity-check code (LDPC) is typically adopted as an error correction code (ECC) to encode data and provide fault tolerance. To reduce the cost, LDPC with a high code rate is always adopted. However, LDPC will lead to read retry operations when the accessed data are not successfully decoded, and such retry-induced performance degradation is serious, especially for modern high-density flash memory. In this work, a reliability-aware differential ECC (READECC) approach is proposed to reduce redundancy protection and storage cost of LDPC with a low code rate and optimize the read performance. The basic idea is to adopt LDPC with a suitable code rate considering both data access characteristics and flash reliability characteristics. First, hot reads are identified based on the frequency of being accessed. Second, based on the reliability variation characteristics, the life of flash memory is divided into three reliability periods. As the reliability period shifts, the code rate of the LDPC adjusts adaptively to minimize redundancy protection. Third, an adaptive-sized logical page approach is further proposed to support LDPC with strong error correction capability (a low code rate) with a low storage cost. Through careful design and evaluation on 3-D triple-level-cell <sc>nand</small> flash memory, READECC achieves encouraging optimizations with a negligible cost.","PeriodicalId":13425,"journal":{"name":"IEEE Transactions on Very Large Scale Integration (VLSI) Systems","volume":"33 6","pages":"1679-1692"},"PeriodicalIF":2.8,"publicationDate":"2025-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144117384","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Synthesis of Analog and Mixed-Signal Circuits on a Programmable Array","authors":"Ziyi Chen;Ioannis Savidis","doi":"10.1109/TVLSI.2025.3553538","DOIUrl":"https://doi.org/10.1109/TVLSI.2025.3553538","url":null,"abstract":"In this article, a novel field-programmable analog array (FPAA) has been developed for the configurable implementation of various analog circuits. The proposed architecture not only supports system-level reconfiguration but also enables transistor-level programmability. The FPAA is comprised of a <inline-formula> <tex-math>$3times 4$ </tex-math></inline-formula> configurable analog block (CAB) array, with a single configurable logic block (CLB) added to each column to allow for the programming of digital circuits. Passive devices, including programmable capacitors and resistors, and active transistor pairs (TPs), are utilized to implement both continuous-time and discrete-time circuits. A placement algorithm is developed that efficiently maps analog circuits onto the FPAA fabric by finding the optimal vertical and horizontal locations for the assignment of transistors. In addition, to reduce the complexity of placing devices on the fabric, a technique is developed that matches TPs in the same vertical level to predefined topologies in a library. Routers are included to connect devices implemented on the FPAA fabric. The proposed FPAA occupies an area of 4 mm² in a TSMC 65-nm fabrication process. The smaller circuits implemented on the FPAA fabric include a folded-cascode amplifier, a strongArm comparator, a continuous-time integrator, and a switch-capacitor integrator. The larger analog and mixed-signal circuits implemented on the FPAA fabric include a four-stage pipeline analog-to-digital converter (ADC) and a first-order delta-sigma modulator. The programmed folded-cascode amplifier exhibits a tunable gain of 28.3 dB to 34.8 dB and a programmable 3-dB bandwidth of 3.3 MHz to 5.3 MHz. The configured comparator provides a resolution of less than 3 mV when comparing two signals. The implemented first-order delta-sigma modulator operates at a frequency of 15 MHz and provides an effective number of bits (ENOBs) of 6.8 when utilizing an oversampling ratio of <inline-formula> <tex-math>$128times $ </tex-math></inline-formula>. The configured pipeline ADC provides an ENOB of 3.7 for a sampling frequency of 15 MHz.","PeriodicalId":13425,"journal":{"name":"IEEE Transactions on Very Large Scale Integration (VLSI) Systems","volume":"33 7","pages":"1920-1933"},"PeriodicalIF":2.8,"publicationDate":"2025-03-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144519389","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"An Area-Energy-Efficient 64–2048 Point FFT With Approximate Plane-Fitting Complex Multipliers","authors":"Weiwei Shi;Jiasheng Wu;Yida Yuan;Zhihong Mo;Chaoyuan Wu;Jiangwei He","doi":"10.1109/TVLSI.2025.3550470","DOIUrl":"https://doi.org/10.1109/TVLSI.2025.3550470","url":null,"abstract":"As a key component of fast Fourier transform (FFT), the complex multiplier (CM) includes twiddle factor generation and corresponding multiplication. This brief proposes an tailored approach for approximating CM functionality by employing an adapted piecewise-plane-fitting technique, effectively replacing the conventional look-up-table-based twiddle generation and exact multipliers by shift-and-add calculation. Numerical binary calculation analysis and simulations are conducted to achieve an optimal tradeoff among accuracy, circuit complexity, power, and delay. Based on 45-nm CMOS, logic synthesis results demonstrate significant improvements, with area, power, and delay reductions of 64.18%, 64.98%, and 19.77%, respectively. With optimizations on logic structures, the complete design of the 64–2048 point FFT has efficiently adopted the proposed CM with evident improvement. The proposed FFT outperforms other reconfigurable FFT designs in terms of normalized area reduction over 55.53% and normalized energy improvement over 21.51%. In field-programmable gate array (FPGA) implementation, the proposed FFT has significantly more savings compared with the exact FFT. In practice, the approximate FFT output results’ PSNR ranges from 56 to 83 dB with competent accuracy in typical signal processing.","PeriodicalId":13425,"journal":{"name":"IEEE Transactions on Very Large Scale Integration (VLSI) Systems","volume":"33 7","pages":"2034-2038"},"PeriodicalIF":2.8,"publicationDate":"2025-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144519295","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"An Area and Energy-Efficient Systolic Array Accelerator Architecture for Deep Neural Networks Using Stochastic Computing","authors":"Jingwei Zhu;Jingguo Wu;Zongru Yang;Yu Jiang;Yun Chen","doi":"10.1109/TVLSI.2025.3550786","DOIUrl":"https://doi.org/10.1109/TVLSI.2025.3550786","url":null,"abstract":"Deep neural networks (DNNs) are widely used to handle various intelligent tasks. With the increased model size, the DNNs’ hardware accelerators are challenging the higher area overhead and energy consumption. Stochastic computing (SC) has recently been considered for implementing DNNs and reducing hardware consumption. However, many current SC-based DNN accelerators fail to balance accuracy, performance, and resource overhead. In addition, their limited scalability and flexibility restrict their use in edge devices. In this article, we design an area and energy-efficient DNN accelerator architecture using SC. We propose an SC-binary hybrid processing unit with piecewise shift compensation without significant additional hardware overhead increment to improve the SC accuracy. To balance performance and resource overhead, we conduct a design space exploration (DSE) from an overall architectural perspective. An experimental platform with both software and hardware for SC-based DNNs is established. The software simulation results demonstrate that the best accuracy of the designed SC-DNN on the CIFAR-10 is 91.9%, which is 3.2% higher than that of the previous SC-DNN work. The VLSI implementation of the hardware is synthesized using the TSMC 28-nm CMOS process. Results show that compared to the binary computing counterpart, our design achieves <inline-formula> <tex-math>$2.7times $ </tex-math></inline-formula> area efficiency and <inline-formula> <tex-math>$3.4times $ </tex-math></inline-formula> energy efficiency. Compared to other SC-DNN accelerator designs, our design can provide <inline-formula> <tex-math>$5.3times $ </tex-math></inline-formula> area efficiency and <inline-formula> <tex-math>$7.3times $ </tex-math></inline-formula> energy efficiency.","PeriodicalId":13425,"journal":{"name":"IEEE Transactions on Very Large Scale Integration (VLSI) Systems","volume":"33 6","pages":"1582-1595"},"PeriodicalIF":2.8,"publicationDate":"2025-03-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144117276","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}