Yuhan He
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Yuhan He has a total of 20 co-authors. Here are the top co-authors by number of shared publications:
- Lirong Zheng
Fudan University
4 shared publications - Zhuo Zou
Fudan University
4 shared publications - Anqin Xiao
State Key Laboratory of Integrated Chips and Systems, College of Future Information Technology, Fudan University, Shanghai, China
2 shared publications - Fanxi Yang
State Key Laboratory of Integrated Chips and Systems, School of Information Science and Technology, Fudan University, Shanghai, China
2 shared publications - Deyu Wang
State Key Laboratory of Integrated Chips and Systems and the College of Future Information Technology, Fudan University, Shanghai, China
1 shared publication Hanwen Ou
State Key Laboratory of Integrated Chips and Systems, School of Information Science and Technology, Fudan University, Shanghai, China
1 shared publication
Latest Publications
- 11 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal on Emerging and Selected Topics in Circuits and Systems
- Vol. 16 (2) , 348-362
Variation-Tolerant Circuit Design and Online Learning Framework for Memristor-Based Trace-STDP SNN
Spiking neural networks (SNNs) demonstrate significant potential for energy-efficient neuromorphic computing, while trace-based spike-timing dependent plasticity (trace-STDP) serves as a biologically inspired learning rule to improve SNN temporal learning. Memristor-based implementations of trace-STDP SNN leverage device non-linearity to emulate synaptic dynamics, enabling highly energy-efficient computation. However, the mapping from trace-STDP to memristor arrays introduces inherent variations, which significantly degrade network accuracy, posing a critical challenge for the reliable deployment of memristor-based neuromorphic systems. To address this issue, efficient variation-tolerant strategies are essential. However, current approaches do not take the trace-STDP SNN characteristics into consideration, thereby limiting their algorithm-and hardware-level effectiveness. In this work, a variation-aware online learning framework for memristor-based trace-STDP SNN is proposed, with hierarchical simulations integrated across software, ASIC, memristor-array, and memristor-device levels. Through variation analysis experiments, it is observed that variations in near-zero trace values significantly impact network accuracy. To mitigate these effects, a tailored variation-tolerant strategy is developed, together with a configurable hardware clipping unit and an ASIC/memristor co-simulator. The proposed variation-tolerant design achieves up to a 16.97% accuracy improvement, with only a 0.02% extra power overhead and a 0.01% extra area overhead, while a parallel design ensures no additional latency overhead. Integrating the co-simulator into the proposed framework facilitates design space exploration, enhancing the robustness and efficiency of memristor-based neuromorphic computing.
- 11 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal on Emerging and Selected Topics in Circuits and Systems
- Vol. 16 (2) , 348-362
Variation-Tolerant Circuit Design and Online Learning Framework for Memristor-Based Trace-STDP SNN
Spiking neural networks (SNNs) demonstrate significant potential for energy-efficient neuromorphic computing, while trace-based spike-timing dependent plasticity (trace-STDP) serves as a biologically inspired learning rule to improve SNN temporal learning. Memristor-based implementations of trace-STDP SNN leverage device non-linearity to emulate synaptic dynamics, enabling highly energy-efficient computation. However, the mapping from trace-STDP to memristor arrays introduces inherent variations, which significantly degrade network accuracy, posing a critical challenge for the reliable deployment of memristor-based neuromorphic systems. To address this issue, efficient variation-tolerant strategies are essential. However, current approaches do not take the trace-STDP SNN characteristics into consideration, thereby limiting their algorithm-and hardware-level effectiveness. In this work, a variation-aware online learning framework for memristor-based trace-STDP SNN is proposed, with hierarchical simulations integrated across software, ASIC, memristor-array, and memristor-device levels. Through variation analysis experiments, it is observed that variations in near-zero trace values significantly impact network accuracy. To mitigate these effects, a tailored variation-tolerant strategy is developed, together with a configurable hardware clipping unit and an ASIC/memristor co-simulator. The proposed variation-tolerant design achieves up to a 16.97% accuracy improvement, with only a 0.02% extra power overhead and a 0.01% extra area overhead, while a parallel design ensures no additional latency overhead. Integrating the co-simulator into the proposed framework facilitates design space exploration, enhancing the robustness and efficiency of memristor-based neuromorphic computing.
- 28 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Artificial Intelligence
- Vol. 3 (3) , 176-189
A 40-nm Sub-mJ/Transfer HDC-SNN Hybrid Processor Enabling On-Chip Few-Shot Transfer Learning for IoT Applications
This work presents a brain-inspired processor that combines Hyperdimensional Computing (HDC) and Spiking Neural Network (SNN), or HDC-SNN hybrid, to enable on-chip few-shot transfer learning (FSTL) for IoT applications. Such an HDC-SNN hybrid architecture utilizes SNN for efficient feature extraction to suppress the size of the vector dimension, thus mitigating intensive operations and the memory footprint for HDC. It also leverages the few-shot and single-pass learning capability of HDC, eliminating the need for abundant labeled samples and training epochs. This processor utilizes a task-agnostic spiking HDC encoder to generate spatio-temporal feature vectors, followed by an on-chip learning HDC classifier. Furthermore, the implementation of on-chip learning uses the schemes of weight access from destination-neuron with memory bank (WAD-MB) and batch-by-batch gradient computation (BGC), achieving a significant improvement on training latency and cache size. This design is fabricated in a 40-nm CMOS technology with a peak efficiency of sub-mJ/transfer (mJ/class and mJ/patient by on-chip FSTL) at 0.72V and 65MHz, occupying a core area of 1.56 mm${}^{2}$. For typical neuromorphic datasets, this processor achieves a 1.6-35x improvement in training energy per class and a 3.5-52x reduction in training latency per sample, compared to state-of-the-art processors. For patient-specific and class-incremental scenarios, it achieves an accuracy of over 96% in both ECG arrhythmia detection and bearing anomaly detection. Specifically, it requires only 2-minute training samples for a new patient with 0.1 mJ energy, achieving a 4.5× improvement in energy efficiency; it employs 3-shot learning for a new class with only 0.3 mJ energy, reducing training epochs by over 100×.
- 28 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Artificial Intelligence
- Vol. 3 (3) , 176-189
A 40-nm Sub-mJ/Transfer HDC-SNN Hybrid Processor Enabling On-Chip Few-Shot Transfer Learning for IoT Applications
This work presents a brain-inspired processor that combines Hyperdimensional Computing (HDC) and Spiking Neural Network (SNN), or HDC-SNN hybrid, to enable on-chip few-shot transfer learning (FSTL) for IoT applications. Such an HDC-SNN hybrid architecture utilizes SNN for efficient feature extraction to suppress the size of the vector dimension, thus mitigating intensive operations and the memory footprint for HDC. It also leverages the few-shot and single-pass learning capability of HDC, eliminating the need for abundant labeled samples and training epochs. This processor utilizes a task-agnostic spiking HDC encoder to generate spatio-temporal feature vectors, followed by an on-chip learning HDC classifier. Furthermore, the implementation of on-chip learning uses the schemes of weight access from destination-neuron with memory bank (WAD-MB) and batch-by-batch gradient computation (BGC), achieving a significant improvement on training latency and cache size. This design is fabricated in a 40-nm CMOS technology with a peak efficiency of sub-mJ/transfer (mJ/class and mJ/patient by on-chip FSTL) at 0.72V and 65MHz, occupying a core area of 1.56 mm${}^{2}$. For typical neuromorphic datasets, this processor achieves a 1.6-35x improvement in training energy per class and a 3.5-52x reduction in training latency per sample, compared to state-of-the-art processors. For patient-specific and class-incremental scenarios, it achieves an accuracy of over 96% in both ECG arrhythmia detection and bearing anomaly detection. Specifically, it requires only 2-minute training samples for a new patient with 0.1 mJ energy, achieving a 4.5× improvement in energy efficiency; it employs 3-shot learning for a new class with only 0.3 mJ energy, reducing training epochs by over 100×.
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. PP (15498328) , 1-14
PHENICS: A Scalable Neuromorphic FPGA Architecture for Million-Neuron Cortical Simulation With 4.6× Real-Time Acceleration
Traditional brain simulations using CPU/GPU architectures face critical limitations in both speed and scalability, particularly when modeling large-scale biological neural networks. Cortical models exhibit distinctive features, recurrent, random and sparse connectivity, and ultra-high synaptic density that fundamentally differ from the structured and dense architectures of deep neural networks (DNNs). As a result, conventional accelerators suffer from inefficiencies: communication overhead scales linearly with neuron count, and memory architectures exhibit poor utilization efficiency under complex connectivity, severely constraining scalability. To address these challenges, we present PHENICS (Pyramidal Hierarchical Event-driven Neuromorphic Infrastructure for Cortical Simulation), a scalable FPGA-based architecture tailored for large-scale spiking neural network simulations. At the communication level, PHENICS introduces a pyramidal multi-tier on-chip network combined with a Busy-Aware Threshold Adaptation (BATA) routing strategy and a lightweight router design to mitigate network congestion and improve spike transmission efficiency. At the storage level, we employ a multi-level addressing scheme adapted to sparse, irregular synaptic connectivity and leverage high-bandwidth memory (HBM) for efficient large-scale synaptic access. PHENICS is implemented on the Xilinx Alveo U50 system, successfully simulating a one-million-neuron Leaky Integrate-and-Fire (LIF) cortical model with 4 billion synapses, achieving a $4.6\times $ real-time acceleration. It achieves this with only 25% of the memory bandwidth available on GPU platforms, yet delivers a $5\times $ speedup compared to GPU-based simulators. Furthermore, our architecture reduces synaptic storage overhead by $3\times $ through compressed encoding and HBM-backed sparse addressing. Moreover, it exhibits sub-linear communication time growth with increasing neuron counts. These advancements pave the way for real-time simulation of billion-neuron brain-scale models, opening new frontiers in neuroscience and neuromorphic computing.
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. PP (15498328) , 1-14
PHENICS: A Scalable Neuromorphic FPGA Architecture for Million-Neuron Cortical Simulation With 4.6× Real-Time Acceleration
Traditional brain simulations using CPU/GPU architectures face critical limitations in both speed and scalability, particularly when modeling large-scale biological neural networks. Cortical models exhibit distinctive features, recurrent, random and sparse connectivity, and ultra-high synaptic density that fundamentally differ from the structured and dense architectures of deep neural networks (DNNs). As a result, conventional accelerators suffer from inefficiencies: communication overhead scales linearly with neuron count, and memory architectures exhibit poor utilization efficiency under complex connectivity, severely constraining scalability. To address these challenges, we present PHENICS (Pyramidal Hierarchical Event-driven Neuromorphic Infrastructure for Cortical Simulation), a scalable FPGA-based architecture tailored for large-scale spiking neural network simulations. At the communication level, PHENICS introduces a pyramidal multi-tier on-chip network combined with a Busy-Aware Threshold Adaptation (BATA) routing strategy and a lightweight router design to mitigate network congestion and improve spike transmission efficiency. At the storage level, we employ a multi-level addressing scheme adapted to sparse, irregular synaptic connectivity and leverage high-bandwidth memory (HBM) for efficient large-scale synaptic access. PHENICS is implemented on the Xilinx Alveo U50 system, successfully simulating a one-million-neuron Leaky Integrate-and-Fire (LIF) cortical model with 4 billion synapses, achieving a $4.6\times $ real-time acceleration. It achieves this with only 25% of the memory bandwidth available on GPU platforms, yet delivers a $5\times $ speedup compared to GPU-based simulators. Furthermore, our architecture reduces synaptic storage overhead by $3\times $ through compressed encoding and HBM-backed sparse addressing. Moreover, it exhibits sub-linear communication time growth with increasing neuron counts. These advancements pave the way for real-time simulation of billion-neuron brain-scale models, opening new frontiers in neuroscience and neuromorphic computing.
- 2 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 73 (6) , 4232-4245
CAMPRO: A CAM-Based Processing-in-Memory Processor for Hyperdimensional Computing
This work introduces CAMPRO, a Content Addressable Memory (CAM)-based Processing-In-Memory (PIM) processor customized for Hyperdimensional Computing (HDC). CAMPRO leverages a 6T Split Word Lines (SWL) cell structure for its CAM, enabling efficient column-wise search for ultra-wide Hypervector (HV) storage and an optimized associative PIM architecture tailored to HDC operations, significantly enhancing energy efficiency. The four key operators of HDC, binding, bundling, permutation, and similarity, are mapped to the proposed architecture. CAMPRO enhances operational parallelism via approximate bundling and employs a hierarchical permutation method to mitigate the gap in flexible shift support within CAM-based PIM architecture. The fine grained pipelined operations boost processing efficiency and dynamically reclaim memory space to support larger models. The Two-Phase Bit Pruning (TPBP) strategy prunes redundant bits in class HVs across two computing stages to eliminate unnecessary computations, reducing operation counts by 73.6% and energy consumption by 65.2% while maintaining query precision. Simulated in a 22 nm CMOS process, CAMPRO occupies 1.13 mm2 and consumes 0.99 mW at 200 MHz. CAMPRO demonstrates robust versatility and scalability across five datasets, including MUTAG, CIFAR10, MNIST, language classification, and EMG gesture recognition, using diverse encoding schemes. It achieves excellent energy efficiency and low latency from small to large-scale datasets. In language classification, it reduces inference energy by 99.2% compared to the similiar work. For EMG gesture recognition, it improves training and inference energy efficiency by 2.6x and 6.7x, and reduce inference latency by 73x compared to related works. On MNIST, it enhances energy efficiency by 11.3x and latency by 1.6x to the prior work, making it an efficient Artificial Intelligence of Things (AIoT) solution.
- 2 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 73 (6) , 4232-4245
CAMPRO: A CAM-Based Processing-in-Memory Processor for Hyperdimensional Computing
This work introduces CAMPRO, a Content Addressable Memory (CAM)-based Processing-In-Memory (PIM) processor customized for Hyperdimensional Computing (HDC). CAMPRO leverages a 6T Split Word Lines (SWL) cell structure for its CAM, enabling efficient column-wise search for ultra-wide Hypervector (HV) storage and an optimized associative PIM architecture tailored to HDC operations, significantly enhancing energy efficiency. The four key operators of HDC, binding, bundling, permutation, and similarity, are mapped to the proposed architecture. CAMPRO enhances operational parallelism via approximate bundling and employs a hierarchical permutation method to mitigate the gap in flexible shift support within CAM-based PIM architecture. The fine grained pipelined operations boost processing efficiency and dynamically reclaim memory space to support larger models. The Two-Phase Bit Pruning (TPBP) strategy prunes redundant bits in class HVs across two computing stages to eliminate unnecessary computations, reducing operation counts by 73.6% and energy consumption by 65.2% while maintaining query precision. Simulated in a 22 nm CMOS process, CAMPRO occupies 1.13 mm2 and consumes 0.99 mW at 200 MHz. CAMPRO demonstrates robust versatility and scalability across five datasets, including MUTAG, CIFAR10, MNIST, language classification, and EMG gesture recognition, using diverse encoding schemes. It achieves excellent energy efficiency and low latency from small to large-scale datasets. In language classification, it reduces inference energy by 99.2% compared to the similiar work. For EMG gesture recognition, it improves training and inference energy efficiency by 2.6x and 6.7x, and reduce inference latency by 73x compared to related works. On MNIST, it enhances energy efficiency by 11.3x and latency by 1.6x to the prior work, making it an efficient Artificial Intelligence of Things (AIoT) solution.