Chixiao Chen
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
Chixiao Chen has a total of 184 co-authors. Here are the top co-authors by number of shared publications:
- Qi Liu
Fudan University
20 shared publications - Haozhe Zhu
Fudan University
17 shared publications Ming Liu
Fudan University
8 shared publications- Siqi He
Fudan University
8 shared publications - Xiaoyang Zeng
Fudan University
7 shared publications - Mengjie Li
State Key Laboratory of Integrated Chips and Systems, College of Integrated Circuits and Micro-Nano Electronics, Fudan University, Shanghai, China
6 shared publications
Latest Publications
- Open Access
- 3 November 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
- Vol. 32 (1) , 1-25
CHARIOT: A Communication-Aware Exploration Framework for 2.5D/3D Silicon Interposer Design
Interposer design is crucial for optimizing inter-chiplet communication efficiency, performance, and energy consumption in 2.5D/3D integrated systems. To facilitate performance analysis and design space exploration (DSE) of interposers, this work presents CHARIOT, a fast, comprehensive exploration framework that establishes a systematic methodology for analyzing and exploring the design space of both passive and active interposers’ architecture. The design space of interposers is first developed to evaluate system performance. Then a roofline model-based methodology is introduced to analyze the bottleneck of performance. A multi-objective Bayesian optimization-based DSE framework, CHARIOT, is proposed, featuring a scheduling engine tailored to different interposer types and a communication-aware exploration engine for coordinated multi-objective exploration of the design space. Compared to the state-of-the-art (SOTA) chiplet system, the CHARIOT-generated optimal passive and active interposer designs achieve energy efficiency improvement up to 1.33× and 1.96×, as well as performance gains up to 1.49× and 1.75× speed-up. The Pareto frontiers of competitive design choices enable selecting the most suitable option based on the designers’ requirements.
- 3 November 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
- Vol. 32 (1) , 1-25
Open AccessCHARIOT: A Communication-Aware Exploration Framework for 2.5D/3D Silicon Interposer Design
Interposer design is crucial for optimizing inter-chiplet communication efficiency, performance, and energy consumption in 2.5D/3D integrated systems. To facilitate performance analysis and design space exploration (DSE) of interposers, this work presents CHARIOT, a fast, comprehensive exploration framework that establishes a systematic methodology for analyzing and exploring the design space of both passive and active interposers’ architecture. The design space of interposers is first developed to evaluate system performance. Then a roofline model-based methodology is introduced to analyze the bottleneck of performance. A multi-objective Bayesian optimization-based DSE framework, CHARIOT, is proposed, featuring a scheduling engine tailored to different interposer types and a communication-aware exploration engine for coordinated multi-objective exploration of the design space. Compared to the state-of-the-art (SOTA) chiplet system, the CHARIOT-generated optimal passive and active interposer designs achieve energy efficiency improvement up to 1.33× and 1.96×, as well as performance gains up to 1.49× and 1.75× speed-up. The Pareto frontiers of competitive design choices enable selecting the most suitable option based on the designers’ requirements.
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 61 (11) , 6110-6122
THz-TSI: A Terahertz Wireless Through-Silicon Interconnect Module for Advanced 3-D Integration Applications
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 61 (11) , 6110-6122
THz-TSI: A Terahertz Wireless Through-Silicon Interconnect Module for Advanced 3-D Integration Applications
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 45 (11) , 5272-5277
SpikeLet: Empowering Multichiplet Neuromorphic System With Hierarchical Interconnect and Spatial–Temporal Parallelism
As Spiking Neural Networks (SNNs) scale, their neuron counts exceed the capacity of a single chip, necessitating multi-chip integration. Multi-chiplet systems provide dense, high-bandwidth interconnects but suffer from inefficient inter-chip spike transmission, limited parallelism, and the lack of transmission-aware mapping. We present SpikeLet, the first multi-chiplet neuromorphic system tailored for large-scale SNNs. SpikeLet employs a hierarchical transmission interconnect with an extend-AER protocol to aggregate spikes and decouple intra- and inter-chiplet communication, an event-driven spatial–temporal parallel dataflow that enables cross-layer and cross-timestep parallelism, and a communication-aware hierarchical mapping based on cooperative co-evolution to balance computation and communication across chiplets and cores. On a 4×4-chiplet prototype, SpikeLet reduces transmission latency by 74.2%, improves throughput by 36.6×, and boosts energy efficiency by 39.2× over a 2D multi-chip baseline, demonstrating a compact and scalable platform for next-generation neuromorphic computing.
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 45 (11) , 5272-5277
SpikeLet: Empowering Multichiplet Neuromorphic System With Hierarchical Interconnect and Spatial–Temporal Parallelism
As Spiking Neural Networks (SNNs) scale, their neuron counts exceed the capacity of a single chip, necessitating multi-chip integration. Multi-chiplet systems provide dense, high-bandwidth interconnects but suffer from inefficient inter-chip spike transmission, limited parallelism, and the lack of transmission-aware mapping. We present SpikeLet, the first multi-chiplet neuromorphic system tailored for large-scale SNNs. SpikeLet employs a hierarchical transmission interconnect with an extend-AER protocol to aggregate spikes and decouple intra- and inter-chiplet communication, an event-driven spatial–temporal parallel dataflow that enables cross-layer and cross-timestep parallelism, and a communication-aware hierarchical mapping based on cooperative co-evolution to balance computation and communication across chiplets and cores. On a 4×4-chiplet prototype, SpikeLet reduces transmission latency by 74.2%, improves throughput by 36.6×, and boosts energy efficiency by 39.2× over a 2D multi-chip baseline, demonstrating a compact and scalable platform for next-generation neuromorphic computing.
- Open Access
- 29 August 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 23 (3) , 1-25
GUMPIM: Unitary and Malleable Memory for Processing-in-Memory with Guaranteed PIM Pages
DRAM-based Processing-in-Memory (PIM) addresses the “memory wall” by executing computations directly inside main memory. However, memory interleaving and virtual memory limit contiguous data size visible to PIM units, constraining PIM task granularity. Fine-grained PIM tasks incur significant offloading overhead that negates PIM performance benefits. To mitigate this, existing PIM systems drastically isolate PIM memory or disable memory interleaving. These design choices, however, decrease the CPU memory bandwidth and introduce extra data transfer, leading to an additional “system memory wall” that degrades CPU performance and must be resolved to realize PIM’s full potential. In this work, we propose GUMPIM, a PIM system that allows interleaved CPU pages and non-interleaved PIM pages to coexist in a U nitary and M alleable memory space with G uaranteed PIM page allocation. GUMPIM enables zero-copy during PIM task offloading and maintains CPU memory bandwidth while ensuring low PIM offloading overhead. First, we propose a dual-track memory management mechanism consisting of independent page allocation and address translation for CPU and PIM pages. Second, we design GUMPIM interface hardware on PIM-enabled DRAMs to provide a dynamic address mapping for the different data layouts of CPU and PIM pages. Third, we propose a PIM-assisted page migration mechanism that transparently migrates pages while preserving CPU access bandwidth, thereby enabling guaranteed and accelerated PIM page allocation. GUMPIM requires no changes to commodity DRAM standards; all hardware modifications are limited to the DRAM side, ensuring full compatibility with existing CPUs and enabling immediate deployment on current HBMx- and LPDDRx-based PIM platforms. Our results show only <0.1% performance degradation for CPU workloads on GUMPIM, in contrast to the 25.8% degradation on PIM systems with memory interleaving turned off. For PIM workloads, GUMPIM reduces memory allocation and CPU-part computation times by 2.7× and 4.93×, respectively, yielding an end-to-end 2.3× speedup over a state-of-the-art baseline system.
- 29 August 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 23 (3) , 1-25
Open AccessGUMPIM: Unitary and Malleable Memory for Processing-in-Memory with Guaranteed PIM Pages
DRAM-based Processing-in-Memory (PIM) addresses the “memory wall” by executing computations directly inside main memory. However, memory interleaving and virtual memory limit contiguous data size visible to PIM units, constraining PIM task granularity. Fine-grained PIM tasks incur significant offloading overhead that negates PIM performance benefits. To mitigate this, existing PIM systems drastically isolate PIM memory or disable memory interleaving. These design choices, however, decrease the CPU memory bandwidth and introduce extra data transfer, leading to an additional “system memory wall” that degrades CPU performance and must be resolved to realize PIM’s full potential. In this work, we propose GUMPIM, a PIM system that allows interleaved CPU pages and non-interleaved PIM pages to coexist in a U nitary and M alleable memory space with G uaranteed PIM page allocation. GUMPIM enables zero-copy during PIM task offloading and maintains CPU memory bandwidth while ensuring low PIM offloading overhead. First, we propose a dual-track memory management mechanism consisting of independent page allocation and address translation for CPU and PIM pages. Second, we design GUMPIM interface hardware on PIM-enabled DRAMs to provide a dynamic address mapping for the different data layouts of CPU and PIM pages. Third, we propose a PIM-assisted page migration mechanism that transparently migrates pages while preserving CPU access bandwidth, thereby enabling guaranteed and accelerated PIM page allocation. GUMPIM requires no changes to commodity DRAM standards; all hardware modifications are limited to the DRAM side, ensuring full compatibility with existing CPUs and enabling immediate deployment on current HBMx- and LPDDRx-based PIM platforms. Our results show only <0.1% performance degradation for CPU workloads on GUMPIM, in contrast to the 25.8% degradation on PIM systems with memory interleaving turned off. For PIM workloads, GUMPIM reduces memory allocation and CPU-part computation times by 2.7× and 4.93×, respectively, yielding an end-to-end 2.3× speedup over a state-of-the-art baseline system.
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 61 (7) , 3233-3245
A 1.25 pJ/Bit 73 Gb/s/ch Wireless Thru-Silicon Interface for 3-D Integration Using Terahertz Near-Field Coupling
The escalating demands of high-performance computing (HPC) and artificial intelligence (AI) drive the adoption of chiplet technologies, positioning 3-D integration as a key enabler for performance scaling. However, conventional 3-D integration schemes based on through-silicon vias (TSVs) face significant challenges, including high fabrication complexity, yield degradation, and thermal reliability issues. This article presents a novel high-speed terahertz (THz) through-silicon interface for cost-effective 3-D integration. By eliminating TSVs, the proposed architecture substantially simplifies fabrication, reduces manufacturing costs, and improves system flexibility. Leveraging near-field THz coupling, the interface achieves high bandwidth and a compact form factor while maintaining competitive energy efficiency and interconnect density. A prototype with two independent data links was implemented in a standard 28-nm CMOS process. The QPSK link demonstrates a data rate of 73 Gb/s with an energy efficiency of 1.25 pJ/bit, while the PAM4 link achieves 71.4 Gb/s at 1.22 pJ/bit.
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 61 (7) , 3233-3245
A 1.25 pJ/Bit 73 Gb/s/ch Wireless Thru-Silicon Interface for 3-D Integration Using Terahertz Near-Field Coupling
The escalating demands of high-performance computing (HPC) and artificial intelligence (AI) drive the adoption of chiplet technologies, positioning 3-D integration as a key enabler for performance scaling. However, conventional 3-D integration schemes based on through-silicon vias (TSVs) face significant challenges, including high fabrication complexity, yield degradation, and thermal reliability issues. This article presents a novel high-speed terahertz (THz) through-silicon interface for cost-effective 3-D integration. By eliminating TSVs, the proposed architecture substantially simplifies fabrication, reduces manufacturing costs, and improves system flexibility. Leveraging near-field THz coupling, the interface achieves high bandwidth and a compact form factor while maintaining competitive energy efficiency and interconnect density. A prototype with two independent data links was implemented in a standard 28-nm CMOS process. The QPSK link demonstrates a data rate of 73 Gb/s with an energy efficiency of 1.25 pJ/bit, while the PAM4 link achieves 71.4 Gb/s at 1.22 pJ/bit.
- 16 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Solid-State Circuits Letters
- Vol. PP, 1
An 800MS/s 13b 2x TI Pipelined-SAR ADC with Rapid Digital Amplification
This work proposes a rapid digital amplification (RDA) with residue-aware reference, offering an equivalent open-loop (OL) gain enhancement of 25 dB and reducing the inter-stage gain error (ISGE)-induced SNR degradation by 20 dB, with an extra amplification latency of only 200 ps. It is implemented in an 800MS/s 13b two-way time-interleaved (TI) pipelined-SAR ADC. The ADC exhibits only 2.6dB SNDR variation across different temperatures and supply voltages without requiring additional ISGE calibration. The prototype ADC is implemented in a 28-nm CMOS process and consumes 11.44 mW under 800MS/s. The measured signal-to-noise and distortion ratio (SNDR) and spurious free dynamic range (SFDR) with 345MHz input are 64.69 dB and 76.78 dB, respectively, leading to a Schreier figure-of-merit (FoM) of 170.13 dB.
- 16 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Solid-State Circuits Letters
- Vol. PP, 1
An 800MS/s 13b 2x TI Pipelined-SAR ADC with Rapid Digital Amplification
This work proposes a rapid digital amplification (RDA) with residue-aware reference, offering an equivalent open-loop (OL) gain enhancement of 25 dB and reducing the inter-stage gain error (ISGE)-induced SNR degradation by 20 dB, with an extra amplification latency of only 200 ps. It is implemented in an 800MS/s 13b two-way time-interleaved (TI) pipelined-SAR ADC. The ADC exhibits only 2.6dB SNDR variation across different temperatures and supply voltages without requiring additional ISGE calibration. The prototype ADC is implemented in a 28-nm CMOS process and consumes 11.44 mW under 800MS/s. The measured signal-to-noise and distortion ratio (SNDR) and spurious free dynamic range (SFDR) with 345MHz input are 64.69 dB and 76.78 dB, respectively, leading to a Schreier figure-of-merit (FoM) of 170.13 dB.
- 3 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. PP (15498328) , 1-13
A 1024-Ch 583-nW/Ch Spike-Sorting SoC With Sparsity-Aware Spike Detection Scratchpad and Ultra-Low-Leakage Dual-Voltage 5T-SRAM for 16K-Template Clustering
This paper presents an energy-efficient spike-sorting system-on-chip (SoC) designed for closed-loop brain-computer interfaces of massive probing channels. The design first incorporates a sparsity/similarity-aware spike detection scratchpad, leveraging a bit-wise differential encoder and zero-friendly read-out circuits, reducing the dynamic power consumption of spike detection by 77.7%. To mitigate static power dissipation, it also introduces an ultra-low-leakage dual-voltage 5T-SRAM array with level-shifter embedded sense amplifiers, achieving an 82.2% leakage power reduction of neural signal buffering by applying half $V_{DD}$ on SRAM cells. Additionally, a memory hierarchy architecture combining on-chip SRAM and off-chip FeRAM, along with a firing-rate-based Osort for cluster template management, minimizes off-chip memory access to only 9.7% with a latency of $11.7\mu $ s for 1024-channel spike sorting. A silicon prototype is fabricated in 28-nm CMOS technology, which achieves a power consumption of 583nW/channel and an area consumption of 0.0012mm2/channel. The chip supports real-time spike sorting with up to 16K templates, $21.3\times $ greater than the state-of-the-art spike-sorting processor.
- 3 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. PP (15498328) , 1-13
A 1024-Ch 583-nW/Ch Spike-Sorting SoC With Sparsity-Aware Spike Detection Scratchpad and Ultra-Low-Leakage Dual-Voltage 5T-SRAM for 16K-Template Clustering
This paper presents an energy-efficient spike-sorting system-on-chip (SoC) designed for closed-loop brain-computer interfaces of massive probing channels. The design first incorporates a sparsity/similarity-aware spike detection scratchpad, leveraging a bit-wise differential encoder and zero-friendly read-out circuits, reducing the dynamic power consumption of spike detection by 77.7%. To mitigate static power dissipation, it also introduces an ultra-low-leakage dual-voltage 5T-SRAM array with level-shifter embedded sense amplifiers, achieving an 82.2% leakage power reduction of neural signal buffering by applying half $V_{DD}$ on SRAM cells. Additionally, a memory hierarchy architecture combining on-chip SRAM and off-chip FeRAM, along with a firing-rate-based Osort for cluster template management, minimizes off-chip memory access to only 9.7% with a latency of $11.7\mu $ s for 1024-channel spike sorting. A silicon prototype is fabricated in 28-nm CMOS technology, which achieves a power consumption of 583nW/channel and an area consumption of 0.0012mm2/channel. The chip supports real-time spike sorting with up to 16K templates, $21.3\times $ greater than the state-of-the-art spike-sorting processor.
- 3 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) , 469-482
Scalable Sparse Transformer Accelerator With In-Memory Butterfly Zero Skipper and Local Attention Reusable Engine for Semi-Structured-Pruned NN
Transformer-based large language models (LLMs) have achieved unprecedented advances across diverse AI tasks. However, their execution remains power-hungry, primarily due to the rapidly growing model size and the quadratic complexity of the self-attention. In-memory computing (IMC) has emerged as a promising solution for accelerating LLMs, delivering high storage density, wide memory bandwidth, and intrinsic compute parallelism. To further improve energy efficiency, this paper addresses LLM inference by leveraging irregular sparse computing in digital SRAM-based IMCs. The proposed butterfly-network-based feed-forward IMC engine enables irregular zero-weight skipping while sustaining high compute utilization. In parallel, a local-attention-reusable IMC engine supports variable sparse attention spans and eliminates redundant self-attention operations. Finally, we implement a four-chip system to scale the parallel performance of the IMC chip. Extensive experiments are deployed on the system, including BERT, GPT to Llama models on 11 AI tasks. The results show that the fabricated accelerator in 28nm CMOS achieves a peak throughput of 186.9 tokens/s and an energy cost of 0.21 J/token, yielding 4.2× speedup and 27.6× energy efficiency over the A100 GPU. Compared to the existing LLM-specific accelerators, the proposed accelerator further achieves 1.2× to 16.9× energy efficiency.
- 3 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) , 469-482
Scalable Sparse Transformer Accelerator With In-Memory Butterfly Zero Skipper and Local Attention Reusable Engine for Semi-Structured-Pruned NN
Transformer-based large language models (LLMs) have achieved unprecedented advances across diverse AI tasks. However, their execution remains power-hungry, primarily due to the rapidly growing model size and the quadratic complexity of the self-attention. In-memory computing (IMC) has emerged as a promising solution for accelerating LLMs, delivering high storage density, wide memory bandwidth, and intrinsic compute parallelism. To further improve energy efficiency, this paper addresses LLM inference by leveraging irregular sparse computing in digital SRAM-based IMCs. The proposed butterfly-network-based feed-forward IMC engine enables irregular zero-weight skipping while sustaining high compute utilization. In parallel, a local-attention-reusable IMC engine supports variable sparse attention spans and eliminates redundant self-attention operations. Finally, we implement a four-chip system to scale the parallel performance of the IMC chip. Extensive experiments are deployed on the system, including BERT, GPT to Llama models on 11 AI tasks. The results show that the fabricated accelerator in 28nm CMOS achieves a peak throughput of 186.9 tokens/s and an energy cost of 0.21 J/token, yielding 4.2× speedup and 27.6× energy efficiency over the A100 GPU. Compared to the existing LLM-specific accelerators, the proposed accelerator further achieves 1.2× to 16.9× energy efficiency.