Tao Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Tao Zhang has a total of 22 co-authors. Here are the top co-authors by number of shared publications:
- Yuan Wang
Peking University
5 shared publications - Yi Zhong
Key Laboratory of Microelectronic Devices and Circuits, MPW Center, School of Integrated Circuits, Peking University, Beijing, China
4 shared publications Jinhao Ruan
Key Laboratory of Microelectronic Devices and Circuits, MPW Center, School of Integrated Circuits, Peking University, Beijing, China
3 shared publications- Youming Yang
Peking University
3 shared publications Zhaotong Zhang
Key Laboratory of Microelectronic Devices and Circuits, MPW Center, School of Integrated Circuits, Peking University, Beijing, China
3 shared publicationsYipeng Gao
Key Laboratory of Microelectronic Devices and Circuits, MPW Center, School of Integrated Circuits, Peking University, Beijing, China
2 shared publications
Latest Publications
- 26 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems II: Express Briefs
- Vol. 73 (4) , 468-472
EA-HWP: An Efficient CNN/SNN Accelerator With Hybrid Weight Precision
Due to the low power consumption and event-driven characteristics, spiking neural networks (SNNs) have garnered significant interest in recent years. In comparison, convolutional neural networks (CNNs) achieve state-of-the-art results in image classification tasks by virtue of their local perception and hierarchical learning characteristics. The fusion computing of SNN and CNN represents a promising direction that can fully leverage the low-power advantages of SNN and the high-precision advantages of CNN. In this breif, a 28-nm efficient heterogeneous CNN/SNN accelerator, named EA-HWP is proposed. EA-HWP employs a heterogeneous distributed design for the efficient on-chip implementation of CNN computation, CNN-SNN conversion, and SNN computation. Meanwhile, based on the characteristics of weight distribution, dense-and-sparse computing architecture is proposed for hybrid weight precision computation. Fabricated in 28-nm CMOS process, EA-HWP can run at a maximum of 1152MHz for CNN and 792MHz for SNN. When the supply voltage is 0.625V, EA-HWP can achieve energy efficiency of 0.14pJ/OP for CNN and 0.26pJ/SOP for SNN, respectively.
- 26 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems II: Express Briefs
- Vol. 73 (4) , 468-472
EA-HWP: An Efficient CNN/SNN Accelerator With Hybrid Weight Precision
Due to the low power consumption and event-driven characteristics, spiking neural networks (SNNs) have garnered significant interest in recent years. In comparison, convolutional neural networks (CNNs) achieve state-of-the-art results in image classification tasks by virtue of their local perception and hierarchical learning characteristics. The fusion computing of SNN and CNN represents a promising direction that can fully leverage the low-power advantages of SNN and the high-precision advantages of CNN. In this breif, a 28-nm efficient heterogeneous CNN/SNN accelerator, named EA-HWP is proposed. EA-HWP employs a heterogeneous distributed design for the efficient on-chip implementation of CNN computation, CNN-SNN conversion, and SNN computation. Meanwhile, based on the characteristics of weight distribution, dense-and-sparse computing architecture is proposed for hybrid weight precision computation. Fabricated in 28-nm CMOS process, EA-HWP can run at a maximum of 1152MHz for CNN and 792MHz for SNN. When the supply voltage is 0.625V, EA-HWP can achieve energy efficiency of 0.14pJ/OP for CNN and 0.26pJ/SOP for SNN, respectively.
- 21 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 61 (5) , 2295-2308
SKADI: A 28-nm Complete K-SAT Solver Featuring Bidirectional In-Memory Deduction and Incremental Updating
Boolean satisfiability (K-SAT) is a canonical problem at the core of many electronic design automation (EDA) and artificial intelligence (AI) tasks. Due to its NP-complete characteristic, solving K-SAT ($K \geq 3$ ) problems on von Neumann architectures incurs substantial energy and latency overhead. Recent ASIC K-SAT solvers have significantly improved energy efficiency, but these implementations are inherently incomplete. While these solvers have demonstrated the capability to solve satisfiable (SAT) problems, no evidence has been provided for proving unsatisfiability. This limitation severely restricts their applicability, as real-life scenarios mostly demand definitive SAT/unsatisfiable (UNSAT) outcomes. To address this gap, this work presents a complete K-SAT solver based on the DPLL framework that can determine whether the given formula is SAT or UNSAT. The proposed architecture incorporates a dual-path SRAM-based macro and a position-encoded counter (PEC) to perform highly parallel, bidirectional clause-variable deduction. Additionally, an incremental updating technique is employed for conflict-aware backtracking and search space pruning. Fabricated in 28-nm CMOS, the proposed solver achieves 100% solvability across all SAT and UNSAT benchmarks, with average solution times of$17.1~\mu $ s (uf50-218) and$42.1 ~\mu $ s (uuf50-218) at 0.9-V supply voltage. Corresponding energy consumptions are 58.0 and 142.8 nJ, respectively. As the first ASIC complete K-SAT solver, it enables deployment in applications where deterministically asserting SAT/UNSAT is essential.- 21 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 61 (5) , 2295-2308
SKADI: A 28-nm Complete K-SAT Solver Featuring Bidirectional In-Memory Deduction and Incremental Updating
Boolean satisfiability (K-SAT) is a canonical problem at the core of many electronic design automation (EDA) and artificial intelligence (AI) tasks. Due to its NP-complete characteristic, solving K-SAT ($K \geq 3$ ) problems on von Neumann architectures incurs substantial energy and latency overhead. Recent ASIC K-SAT solvers have significantly improved energy efficiency, but these implementations are inherently incomplete. While these solvers have demonstrated the capability to solve satisfiable (SAT) problems, no evidence has been provided for proving unsatisfiability. This limitation severely restricts their applicability, as real-life scenarios mostly demand definitive SAT/unsatisfiable (UNSAT) outcomes. To address this gap, this work presents a complete K-SAT solver based on the DPLL framework that can determine whether the given formula is SAT or UNSAT. The proposed architecture incorporates a dual-path SRAM-based macro and a position-encoded counter (PEC) to perform highly parallel, bidirectional clause-variable deduction. Additionally, an incremental updating technique is employed for conflict-aware backtracking and search space pruning. Fabricated in 28-nm CMOS, the proposed solver achieves 100% solvability across all SAT and UNSAT benchmarks, with average solution times of$17.1~\mu $ s (uf50-218) and$42.1 ~\mu $ s (uuf50-218) at 0.9-V supply voltage. Corresponding energy consumptions are 58.0 and 142.8 nJ, respectively. As the first ASIC complete K-SAT solver, it enables deployment in applications where deterministically asserting SAT/UNSAT is essential.- 23 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems II: Express Briefs
- Vol. 72 (8) , 1088-1092
UniPRE: An SNN-ANN Accelerator With Unified Max-Pooling Prediction and Redundancy Elimination
The integration of Spiking Neural Networks (SNNs) and Artificial Neural Networks (ANNs) for specific tasks has attracted considerable interest due to their potential for high energy efficiency and accuracy. In SNN-ANN fused hardware, many works focus on neuron-level fusion of operators. Though some have explored optimizations at the dataflow level, they are restricted to only one kind of networks. This paper introduces a dataflow-level unified predicting method to eliminate redundant computations resulted from max-pooling operations for both SNN and ANN by exploiting Channel-wise Importance (CI). An accelerator with online sorting of Channel-wise Importance (CI) to support this optimization is also proposed, named as UniPRE. Results show that UniPRE reduces 44.77% and 31.85% overall computations with negligible accuracy loss for SNN and ANN using 37.5% channels for prediction. Implemented in the standard 28-nm CMOS technology, UniPRE can reach an energy efficiency of 19.32 TSOPS/W and 4.26 TOPS/W, with an area efficiency of 370.10 GSOPS/mm2 and 92.52 GOPS/mm2 for SNN and ANN paradigms of 8-bit weight precision, respectively. In layer-wise evaluation of real networks, up to 1.79× energy reduction is achieved with 25% channels used for prediction.
- 23 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems II: Express Briefs
- Vol. 72 (8) , 1088-1092
UniPRE: An SNN-ANN Accelerator With Unified Max-Pooling Prediction and Redundancy Elimination
The integration of Spiking Neural Networks (SNNs) and Artificial Neural Networks (ANNs) for specific tasks has attracted considerable interest due to their potential for high energy efficiency and accuracy. In SNN-ANN fused hardware, many works focus on neuron-level fusion of operators. Though some have explored optimizations at the dataflow level, they are restricted to only one kind of networks. This paper introduces a dataflow-level unified predicting method to eliminate redundant computations resulted from max-pooling operations for both SNN and ANN by exploiting Channel-wise Importance (CI). An accelerator with online sorting of Channel-wise Importance (CI) to support this optimization is also proposed, named as UniPRE. Results show that UniPRE reduces 44.77% and 31.85% overall computations with negligible accuracy loss for SNN and ANN using 37.5% channels for prediction. Implemented in the standard 28-nm CMOS technology, UniPRE can reach an energy efficiency of 19.32 TSOPS/W and 4.26 TOPS/W, with an area efficiency of 370.10 GSOPS/mm2 and 92.52 GOPS/mm2 for SNN and ANN paradigms of 8-bit weight precision, respectively. In layer-wise evaluation of real networks, up to 1.79× energy reduction is achieved with 25% channels used for prediction.