Jun Tao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Jun Tao has a total of 127 co-authors. Here are the top co-authors by number of shared publications:
- Xuan Zeng
Fudan University
15 shared publications - Xin Li
Data Science Research Center, Duke Kunshan University, Kunshan, Soochow, Jiangsu, China
8 shared publications - Dian Zhou
Microelectronics Department, State Key Laboratory of Integrated Chips and Systems, Fudan University, Shanghai, China
7 shared publications Xin Li
Digital Innovation Research Center, Duke Kunshan University, Kunshan, Jiangsu, China
5 shared publications- Yangfeng Su
School of Mathematical Sciences Fudan University 200433 Shanghai China
4 shared publications - Zhengqi Gao
Massachusetts Institute of Technology
4 shared publications
Latest Publications
- Open Access
- 22 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 392-397
BitPair: An Efficient 2-Bit Serial Precision-Scalable Accelerator for GEMM in Deep Neural Networks
As deep neural networks grow in size, their computational and energy demands pose significant challenges for efficient hardware acceleration. Quantization reduces storage, data movement, and computation costs by lowering numerical precision. While 8-bit integer (INT8) is widely adopted, lower precisions such as INT6 and INT4 can also maintain acceptable accuracy with specialized quantization techniques, motivating accelerators that support multiple precisions. This paper presents BitPair, a precision-scalable accelerator that employs a 2-bit serial dataflow to perform general matrix multiplications (GEMM). BitPair supports signed and unsigned integer operations at 2/4/6/8-bit precision for both input matrices. Implemented in Verilog and synthesized in TSMC 12nm technology, BitPair is compared with two prior precision-scalable architectures, Loom and BitShare. Evaluation shows that BitPair achieves up to 1.63 × lower bandwidth, 1.19 × smaller area, and 1.14 × lower power compared with Loom and BitShare.
- 22 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 392-397
Open AccessBitPair: An Efficient 2-Bit Serial Precision-Scalable Accelerator for GEMM in Deep Neural Networks
As deep neural networks grow in size, their computational and energy demands pose significant challenges for efficient hardware acceleration. Quantization reduces storage, data movement, and computation costs by lowering numerical precision. While 8-bit integer (INT8) is widely adopted, lower precisions such as INT6 and INT4 can also maintain acceptable accuracy with specialized quantization techniques, motivating accelerators that support multiple precisions. This paper presents BitPair, a precision-scalable accelerator that employs a 2-bit serial dataflow to perform general matrix multiplications (GEMM). BitPair supports signed and unsigned integer operations at 2/4/6/8-bit precision for both input matrices. Implemented in Verilog and synthesized in TSMC 12nm technology, BitPair is compared with two prior precision-scalable architectures, Loom and BitShare. Evaluation shows that BitPair achieves up to 1.63 × lower bandwidth, 1.19 × smaller area, and 1.14 × lower power compared with Loom and BitShare.
- 2 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. PP  (02780070) , 1
Efficient Zero-shot Wafer Classification Method based on the Large Multi-modal Model
Wafer map classification is essential for detecting spatial defect patterns in the integrated circuit manufacturing process, thereby enhancing yield and reducing fabrication costs. However, existing approaches, whether supervised or unsupervised, often suffer from limited data, high annotation costs, and poor adaptability to varying production environments. In this paper, we propose a novel zero-shot binary classification method based on a large multi-modal model (LMM), which eliminates the need for labeled training data. The proposed method utilizes visual prompts to highlight key defect regions and textual prompts to describe normal and abnormal patterns in natural language. By integrating visual and textual cues, the LMM effectively distinguishes between normal and abnormal wafer maps. Experimental results demonstrate that our method achieves competitive classification accuracy without model training or fine-tuning, offering a practical and scalable solution for data-scarce and high-variability IC manufacturing scenarios.
- 2 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. PP  (02780070) , 1
Efficient Zero-shot Wafer Classification Method based on the Large Multi-modal Model
Wafer map classification is essential for detecting spatial defect patterns in the integrated circuit manufacturing process, thereby enhancing yield and reducing fabrication costs. However, existing approaches, whether supervised or unsupervised, often suffer from limited data, high annotation costs, and poor adaptability to varying production environments. In this paper, we propose a novel zero-shot binary classification method based on a large multi-modal model (LMM), which eliminates the need for labeled training data. The proposed method utilizes visual prompts to highlight key defect regions and textual prompts to describe normal and abnormal patterns in natural language. By integrating visual and textual cues, the LMM effectively distinguishes between normal and abnormal wafer maps. Experimental results demonstrate that our method achieves competitive classification accuracy without model training or fine-tuning, offering a practical and scalable solution for data-scarce and high-variability IC manufacturing scenarios.
- 4 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 61  (1) , 90-102
SHINSAI: A 586 mm 2 Reusable Active TSV Interposer With Programmable Interconnect Fabric and 512 Mb Underdeck Memory
This article presents SHINSAI—a 586 mm2 reusable active through-silicon via (TSV) interposer addressing key challenges in multi-chiplet integration (MCI) architectures. While active interposers overcome fundamental limitations of passive counterparts by integrating functional circuitry, existing solutions face three critical constraints: 1) non-recurring engineering (NRE) costs from application-specific interposers negating chiplet reuse benefits; 2) memory capacity limitations restricting compute scalability; and 3) inflexible interconnects unable to support diverse communication patterns. SHINSAI introduces three key innovations: a heterogeneous dual-layer network-on-active-interposer (NoAI) combining packet/circuit-switched (CS) paradigms to reduce communication latency; a programmable 2.5-D H-Link fabric enabling protocol-agnostic$\mu $ bump-to-$\mu $ bump routing across arbitrary topologies; and a reconfigurable 3-D vertical link (V-Link) architecture integrating 512 Mb underdeck SRAM with adaptive bandwidth allocation. SHINSAI is implemented in 28-nm CMOS with TSV-middle technology, with a$\mu $ bump-pitch of$40~\mu $ m. A SHINSAI-based MCI prototype system with 16 chiplets is also assembled.- 4 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 61  (1) , 90-102
SHINSAI: A 586 mm 2 Reusable Active TSV Interposer With Programmable Interconnect Fabric and 512 Mb Underdeck Memory
This article presents SHINSAI—a 586 mm2 reusable active through-silicon via (TSV) interposer addressing key challenges in multi-chiplet integration (MCI) architectures. While active interposers overcome fundamental limitations of passive counterparts by integrating functional circuitry, existing solutions face three critical constraints: 1) non-recurring engineering (NRE) costs from application-specific interposers negating chiplet reuse benefits; 2) memory capacity limitations restricting compute scalability; and 3) inflexible interconnects unable to support diverse communication patterns. SHINSAI introduces three key innovations: a heterogeneous dual-layer network-on-active-interposer (NoAI) combining packet/circuit-switched (CS) paradigms to reduce communication latency; a programmable 2.5-D H-Link fabric enabling protocol-agnostic$\mu $ bump-to-$\mu $ bump routing across arbitrary topologies; and a reconfigurable 3-D vertical link (V-Link) architecture integrating 512 Mb underdeck SRAM with adaptive bandwidth allocation. SHINSAI is implemented in 28-nm CMOS with TSV-middle technology, with a$\mu $ bump-pitch of$40~\mu $ m. A SHINSAI-based MCI prototype system with 16 chiplets is also assembled.- 25 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 33  (12) , 3369-3382
Coordinating Binary Trait: Accurate and Lightweight Runtime On-Chip Power Meter Design
As heterogeneous platforms scale, power management issues grow increasingly critical. On-chip power awareness is fundamental to effective power management on such platforms. However, the integration of diverse components leads to higher power consumption and more complex power patterns, complicating the perception of power distribution. To address this challenge, this article presents COordinating BInary Trait (COBIT), which focuses on the design and optimization of on-chip power meters (OPMs) for heterogeneous design, enabling power management in large-scale systems. COBIT extracts essential bitwise wires as features, learns a boosting model based on binary trees, and implements the OPM for runtime power prediction. Additionally, multiobjective optimization algorithms are employed in the design space exploration of the power meter to ensure an accurate and low-overhead integration. Experimental results demonstrate the capabilities of COBIT on heterogeneous hardware. Using only 0.028% bitwise power proxies of all RTL wires in the NVIDIA deep learning accelerator (NVDLA) and 0.036% in the Berkeley out-of-order machine (BOOM) processor core, the per-cycle models achieve a mean absolute percentage error (MAPE) of 1.69% and 2.49%, respectively. Regarding the success rate of predictions on power peaks in NVDLA, COBIT achieves a maximum performance improvement of 12.93% and an average improvement of 8.11% over existing methods. Moreover, the gate-area/power overhead of our OPM on BOOM and NVDLA is 1.25%/0.217% and 0.49%/0.097%, respectively, while performing per-cycle power prediction in just three cycles. Unlike previous approaches, which struggle with balancing accuracy and efficiency, COBIT effectively addresses both challenges, delivering unprecedented benefits for large-scale systems.
- 25 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 33  (12) , 3369-3382
Coordinating Binary Trait: Accurate and Lightweight Runtime On-Chip Power Meter Design
As heterogeneous platforms scale, power management issues grow increasingly critical. On-chip power awareness is fundamental to effective power management on such platforms. However, the integration of diverse components leads to higher power consumption and more complex power patterns, complicating the perception of power distribution. To address this challenge, this article presents COordinating BInary Trait (COBIT), which focuses on the design and optimization of on-chip power meters (OPMs) for heterogeneous design, enabling power management in large-scale systems. COBIT extracts essential bitwise wires as features, learns a boosting model based on binary trees, and implements the OPM for runtime power prediction. Additionally, multiobjective optimization algorithms are employed in the design space exploration of the power meter to ensure an accurate and low-overhead integration. Experimental results demonstrate the capabilities of COBIT on heterogeneous hardware. Using only 0.028% bitwise power proxies of all RTL wires in the NVIDIA deep learning accelerator (NVDLA) and 0.036% in the Berkeley out-of-order machine (BOOM) processor core, the per-cycle models achieve a mean absolute percentage error (MAPE) of 1.69% and 2.49%, respectively. Regarding the success rate of predictions on power peaks in NVDLA, COBIT achieves a maximum performance improvement of 12.93% and an average improvement of 8.11% over existing methods. Moreover, the gate-area/power overhead of our OPM on BOOM and NVDLA is 1.25%/0.217% and 0.49%/0.097%, respectively, while performing per-cycle power prediction in just three cycles. Unlike previous approaches, which struggle with balancing accuracy and efficiency, COBIT effectively addresses both challenges, delivering unprecedented benefits for large-scale systems.
- 25 July 2025
- Published by Elsevier in Integration
- Vol. 105
Robust analog/RF circuit design via Cycle-Consistent Generative Adversarial Networks
- 25 July 2025
- Published by Elsevier in Integration
- Vol. 105
Robust analog/RF circuit design via Cycle-Consistent Generative Adversarial Networks
- 26 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 44  (3) , 1199-1203
Efficient Design Optimization for Diffractive Deep Neural Networks
Since diffractive deep neural network (D2NN) provides a full optical solution to implement deep neural networks (DNNs), it offers ultrafast operation speed and virtually unlimited bandwidth, yielding an alternative-yet-competitive approach for computer-based neural networks. A D2NN is composed of several 3D-printed phase masks as hidden layers and a number of optical detectors at the output. To enable automatic and efficient design of D2NNs, we propose an iterative optimization method to determine the optimal design parameters of D2NNs. During each iteration step, we first optimize the physical parameters for masks (e.g., thicknesses) while fixing the detector parameters (e.g., locations). Next, we exhaustively search the detector parameters with fixed masks. These two steps are repeated until convergence is reached. Our numerical experiments demonstrate that the proposed optimization algorithm can produce a high-performance D2NN achieving 97% accuracy for recognizing handwritten digits.
- 26 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 44  (3) , 1199-1203
Efficient Design Optimization for Diffractive Deep Neural Networks
Since diffractive deep neural network (D2NN) provides a full optical solution to implement deep neural networks (DNNs), it offers ultrafast operation speed and virtually unlimited bandwidth, yielding an alternative-yet-competitive approach for computer-based neural networks. A D2NN is composed of several 3D-printed phase masks as hidden layers and a number of optical detectors at the output. To enable automatic and efficient design of D2NNs, we propose an iterative optimization method to determine the optimal design parameters of D2NNs. During each iteration step, we first optimize the physical parameters for masks (e.g., thicknesses) while fixing the detector parameters (e.g., locations). Next, we exhaustively search the detector parameters with fixed masks. These two steps are repeated until convergence is reached. Our numerical experiments demonstrate that the proposed optimization algorithm can produce a high-performance D2NN achieving 97% accuracy for recognizing handwritten digits.
- 4 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 71  (9) , 4156-4168
FPIA: Communication-Aware Multi-Chiplet Integration With Field-Programmable Interconnect Fabric on Reusable Silicon Interposer
Silicon interposer re-usage is drawing attention for cost-effective multi-chiplet integrated systems. To address the communication awareness of inter/off-chiplet interconnect, the paper proposes a field-programmable interconnect fabric and develops its corresponding automatic physical integration tool. The tile-based fabric consists of turnout, cross-over boxes and parallel tracks. It features micro-bump-wise connecting flexibility and hardware efficiency. The automation flow performs chiplet location optimization and efficient bump-to-bump routing, supporting multi-lane bus interconnect and miscellaneous external ports. The methodology is validated by 9 different integration scenarios, where the routability is guaranteed when the local resource utilization ratio approaches 94.5%. The data’s maximum interconnect latency is 2.2 ns and the energy consumption is 1.18 pJ/bit at a bitrate of 1 Gbps. The latency consumes 16.5 $\times\sim$ 53.4 $\times$ fewer clock cycles than the state-of-the-art network-on-package-based reusable interposer architectures.
- 4 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 71  (9) , 4156-4168
FPIA: Communication-Aware Multi-Chiplet Integration With Field-Programmable Interconnect Fabric on Reusable Silicon Interposer
Silicon interposer re-usage is drawing attention for cost-effective multi-chiplet integrated systems. To address the communication awareness of inter/off-chiplet interconnect, the paper proposes a field-programmable interconnect fabric and develops its corresponding automatic physical integration tool. The tile-based fabric consists of turnout, cross-over boxes and parallel tracks. It features micro-bump-wise connecting flexibility and hardware efficiency. The automation flow performs chiplet location optimization and efficient bump-to-bump routing, supporting multi-lane bus interconnect and miscellaneous external ports. The methodology is validated by 9 different integration scenarios, where the routability is guaranteed when the local resource utilization ratio approaches 94.5%. The data’s maximum interconnect latency is 2.2 ns and the energy consumption is 1.18 pJ/bit at a bitrate of 1 Gbps. The latency consumes 16.5 $\times\sim$ 53.4 $\times$ fewer clock cycles than the state-of-the-art network-on-package-based reusable interposer architectures.
- Open Access
- 23 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Auto-ISP: An Efficient Real-Time Automatic Hyperparameter Optimization Framework for ISP Hardware System
Image Signal Processor (ISP) is widely used in intelligent edge devices across various scenarios. The intricate and time-consuming tuning process demands substantial expertise. Current AI-based auto-tuning operates discretely offline, relying on predefined scenes with human intervention, leading to inconvenient manipulation, with potentially fatal impacts on downstream tasks in unforeseen scenes. We propose a real-time automatic hyperparameter optimization ISP hardware system to address real-world scenarios. Our design features a tri-step framework and a hardware accelerator, demonstrating superior performance in human and computer vision tasks, even in real-time unforeseen scenes. Experiments showcase its practicality, achieving 1080P@75FPS/240FPS in FPGA/ASIC, respectively.
- 23 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Open AccessAuto-ISP: An Efficient Real-Time Automatic Hyperparameter Optimization Framework for ISP Hardware System
Image Signal Processor (ISP) is widely used in intelligent edge devices across various scenarios. The intricate and time-consuming tuning process demands substantial expertise. Current AI-based auto-tuning operates discretely offline, relying on predefined scenes with human intervention, leading to inconvenient manipulation, with potentially fatal impacts on downstream tasks in unforeseen scenes. We propose a real-time automatic hyperparameter optimization ISP hardware system to address real-world scenarios. Our design features a tri-step framework and a hardware accelerator, demonstrating superior performance in human and computer vision tasks, even in real-time unforeseen scenes. Experiments showcase its practicality, achieving 1080P@75FPS/240FPS in FPGA/ASIC, respectively.
- 19 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems II: Express Briefs
- Vol. 71  (5) , 2644-2648
A Low Valid Throughput Loss LDPC Codec Architecture With Variable Code Rate
Shortening is a rate-compatible technology that can improve LDPC error correction performance by modifying the code rate. However, as the code rate decreases, the throughput of valid data drops rapidly. Reducing the attenuation of valid throughput at low code rates is challenging. For this reason, this brief proposes two contributions. First, a Block Tag-based encoding method is presented with a non-fixed code length, which reduces the latency caused by fixed-length communications. Second, a novel Time Division Multiplexing (TDM) architecture is developed to improve valid throughput, reducing resource utilization compared to conventional multiplexers. Based on the above contributions, the codec is implemented on FPGA and the results show that the valid throughput loss is low in changing from high to low code rates.
- 19 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems II: Express Briefs
- Vol. 71  (5) , 2644-2648
A Low Valid Throughput Loss LDPC Codec Architecture With Variable Code Rate
Shortening is a rate-compatible technology that can improve LDPC error correction performance by modifying the code rate. However, as the code rate decreases, the throughput of valid data drops rapidly. Reducing the attenuation of valid throughput at low code rates is challenging. For this reason, this brief proposes two contributions. First, a Block Tag-based encoding method is presented with a non-fixed code length, which reduces the latency caused by fixed-length communications. Second, a novel Time Division Multiplexing (TDM) architecture is developed to improve valid throughput, reducing resource utilization compared to conventional multiplexers. Based on the above contributions, the codec is implemented on FPGA and the results show that the valid throughput loss is low in changing from high to low code rates.
- Open Access
- 15 November 2023
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
- Vol. 29  (1) , 1-17
Yield Optimization for Analog Circuits over Multiple Corners via Bayesian Neural Networks: Enhancing Circuit Reliability under Environmental Variation
The reliability of circuits is significantly affected by process variations in manufacturing and environmental variation during operation. Current yield optimization algorithms take process variations into consideration to improve circuit reliability. However, the influence of environmental variations (e.g., voltage and temperature variations) is often ignored in current methods because of the high computational cost. In this article, a novel and efficient approach named BNN-BYO is proposed to optimize the yield of analog circuits in multiple environmental corners. First, we use a Bayesian Neural Network (BNN) to simultaneously model the yields and performances of interest in multiple corners efficiently. Next, the multi-corner yield optimization can be performed by embedding BNN into a Bayesian optimization framework. Since the correlation among yields and performances of interest in different corners is implicitly encoded in the BNN model, it provides great modeling capabilities for yields and their uncertainties to improve the efficiency of yield optimization. Our experimental results demonstrate that the proposed method can save up to 45.3% of simulation cost compared to other baseline methods to achieve the same target yield. In addition, for the same simulation cost, our proposed method can find better design points with 3.2% yield improvement.
- 15 November 2023
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
- Vol. 29  (1) , 1-17
Open AccessYield Optimization for Analog Circuits over Multiple Corners via Bayesian Neural Networks: Enhancing Circuit Reliability under Environmental Variation
The reliability of circuits is significantly affected by process variations in manufacturing and environmental variation during operation. Current yield optimization algorithms take process variations into consideration to improve circuit reliability. However, the influence of environmental variations (e.g., voltage and temperature variations) is often ignored in current methods because of the high computational cost. In this article, a novel and efficient approach named BNN-BYO is proposed to optimize the yield of analog circuits in multiple environmental corners. First, we use a Bayesian Neural Network (BNN) to simultaneously model the yields and performances of interest in multiple corners efficiently. Next, the multi-corner yield optimization can be performed by embedding BNN into a Bayesian optimization framework. Since the correlation among yields and performances of interest in different corners is implicitly encoded in the BNN model, it provides great modeling capabilities for yields and their uncertainties to improve the efficiency of yield optimization. Our experimental results demonstrate that the proposed method can save up to 45.3% of simulation cost compared to other baseline methods to achieve the same target yield. In addition, for the same simulation cost, our proposed method can find better design points with 3.2% yield improvement.