Junkang Zhu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Junkang Zhu has a total of 37 co-authors. Here are the top co-authors by number of shared publications:
- Zhengya Zhang
University of Michigan
5 shared publications - Wei Tang
University of Michigan
4 shared publications - Cheng-Hsun Lu
Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI, USA
2 shared publications - Minsik Kim
University of Michigan
2 shared publications Tianyu Wei
University of Michigan
2 shared publications- Allen Chan
Intel Corporation, Santa Clara, CA, USA
1 shared publication
Latest Publications
- Open Access
- 20 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 313-326
HiPER: Hierarchically-Composed Processing for Efficient Robot Learning-Based Control
- 20 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 313-326
Open AccessHiPER: Hierarchically-Composed Processing for Efficient Robot Learning-Based Control
- 21 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 60 (8) , 2995-3008
An 8-bit 20.7 TOPS/W Multilevel Cell ReRAM Macro With ADC-Assisted Bit-Serial Processing
Analog compute in memory (CIM) with multilevel cell (MLC) resistive random access memory (ReRAM) promises highly dense and efficient compute support for machine learning and scientific computing. This article introduces analog to digital converter (ADC)-assisted bit-serial processing for efficient, high-throughput compute. Bit-serial digital to analog converters (DACs) and 8-bit binary-weighted multicycle sampling (BWMCS) ADCs perform analog vector-matrix multiplication (VMM) on MLC-based crossbar arrays. A direct drive ${g}_{m}$ -boosted transimpedance amplifier (TIA) enables high-speed crossbar readout. We present a system on chip (SoC) prototype consisting of four self-contained ReRAM-based CIM macros and a reduced instruction set computer-five (RISC-V) host. The test chip is fabricated in 65 nm CMOS with foundry-integrated MLC ReRAM. We trained LeNet1 for handwritten digit classification and mapped the CNN weights differentially to 3-bit MLC ReRAM across multiple CIM macros. The classification accuracy loss is 1.6% when compared to the quantization-aware trained model. The measured raw and normalized peak efficiencies are 20.7 and 662 TOPS/W, respectively. The compute density is 8.4 TOPS/mm2.
- 21 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 60 (8) , 2995-3008
An 8-bit 20.7 TOPS/W Multilevel Cell ReRAM Macro With ADC-Assisted Bit-Serial Processing
Analog compute in memory (CIM) with multilevel cell (MLC) resistive random access memory (ReRAM) promises highly dense and efficient compute support for machine learning and scientific computing. This article introduces analog to digital converter (ADC)-assisted bit-serial processing for efficient, high-throughput compute. Bit-serial digital to analog converters (DACs) and 8-bit binary-weighted multicycle sampling (BWMCS) ADCs perform analog vector-matrix multiplication (VMM) on MLC-based crossbar arrays. A direct drive ${g}_{m}$ -boosted transimpedance amplifier (TIA) enables high-speed crossbar readout. We present a system on chip (SoC) prototype consisting of four self-contained ReRAM-based CIM macros and a reduced instruction set computer-five (RISC-V) host. The test chip is fabricated in 65 nm CMOS with foundry-integrated MLC ReRAM. We trained LeNet1 for handwritten digit classification and mapped the CNN weights differentially to 3-bit MLC ReRAM across multiple CIM macros. The classification accuracy loss is 1.6% when compared to the quantization-aware trained model. The measured raw and normalized peak efficiencies are 20.7 and 662 TOPS/W, respectively. The compute density is 8.4 TOPS/mm2.
- 27 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 59 (4) , 1235-1245
Arvon: A Heterogeneous System-in-Package Integrating FPGA and DSP Chiplets for Versatile Workload Acceleration
Integrating heterogeneous chiplets in a package presents a promising and cost-effective approach to constructing scalable and flexible systems for accelerating a wide range of workloads. We introduce Arvon that integrates a 14-nm FPGA chiplet with two efficient and densely packed 22-nm DSP chiplets using embedded multidie interconnect bridges (EMIBs). The chiplets are interconnected via a 1.536-Tb/s advanced interface bus (AIB) 1.0 interface and a 7.68-Tb/s AIB 2.0 interface. Arvon is programmable, supporting various workloads from neural network (NN) to communication signal processing. Each DSP chiplet delivers a peak performance of 4.14 TFLOPS in half-precision floating-point while maintaining a power efficiency of 1.8 TFLOPS/W. A compilation procedure is developed to map workloads across the FPGA and DSPs to optimize performance and utilization. Our AIB 2.0 interface implementation using 36- $\mu$ m-pitch microbumps achieves a data transfer rate of 4 Gb/s/pin, with an energy efficiency of 0.10–0.46 pJ/b including the adapter. The bandwidth density reaches 1.024 Tb/s/mm of shoreline and 1.705 Tb/s/mm 2 of area.
- 27 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 59 (4) , 1235-1245
Arvon: A Heterogeneous System-in-Package Integrating FPGA and DSP Chiplets for Versatile Workload Acceleration
Integrating heterogeneous chiplets in a package presents a promising and cost-effective approach to constructing scalable and flexible systems for accelerating a wide range of workloads. We introduce Arvon that integrates a 14-nm FPGA chiplet with two efficient and densely packed 22-nm DSP chiplets using embedded multidie interconnect bridges (EMIBs). The chiplets are interconnected via a 1.536-Tb/s advanced interface bus (AIB) 1.0 interface and a 7.68-Tb/s AIB 2.0 interface. Arvon is programmable, supporting various workloads from neural network (NN) to communication signal processing. Each DSP chiplet delivers a peak performance of 4.14 TFLOPS in half-precision floating-point while maintaining a power efficiency of 1.8 TFLOPS/W. A compilation procedure is developed to map workloads across the FPGA and DSPs to optimize performance and utilization. Our AIB 2.0 interface implementation using 36- $\mu$ m-pitch microbumps achieves a data transfer rate of 4 Gb/s/pin, with an energy efficiency of 0.10–0.46 pJ/b including the adapter. The bandwidth density reaches 1.024 Tb/s/mm of shoreline and 1.705 Tb/s/mm 2 of area.
- 6 May 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 57 (11) , 3490-3502
VOTA: A Heterogeneous Multicore Visual Object Tracking Accelerator Using Correlation Filters
A correlation filter (CF) is a lighter and faster solution for visual object tracking (VOT) compared with a deep neural network (DNN). However, the CF-based trackers introduce diverse computational kernels and require FP16 computations. To address these challenges, we present a VOT accelerator (VOTA), a domain-specific accelerator for CF-based VOT that meets real-time performance at high efficiency, while providing flexibility and programmability. The VOTA encompasses a Winograd convolution core (WINO), a fast Fourier transformation (FFT) core (FFT), and a vector core (VEC) for diverse kernels, integrated in a high-bandwidth star-ring topology. The VOTA’s frame-based instruction set and execution enable a 537 GFLOPS performance, adapt to variances of CF trackers, and reduce the code size. An instruction-chaining mechanism permits inter-core pipelining and improves the hardware utilization up to 84.2%. A 10.2-mm 2 28-nm FP16 system on chip (SoC) prototype incorporating the VOTA, an RISC-V host CPU, and other supportive peripherals is taped out and measured to achieve 2.45 TFLOPS/W at 0.72 V. Running the oriented particle CF (OPCF), a CF tracker enhanced by adaptive boosting and particle filtering, the SoC chip achieves 1157 frames/s ( $640\times480$ frame size) at 0.9 V and 175 MHz, consuming 296 mW.
- 6 May 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 57 (11) , 3490-3502
VOTA: A Heterogeneous Multicore Visual Object Tracking Accelerator Using Correlation Filters
A correlation filter (CF) is a lighter and faster solution for visual object tracking (VOT) compared with a deep neural network (DNN). However, the CF-based trackers introduce diverse computational kernels and require FP16 computations. To address these challenges, we present a VOT accelerator (VOTA), a domain-specific accelerator for CF-based VOT that meets real-time performance at high efficiency, while providing flexibility and programmability. The VOTA encompasses a Winograd convolution core (WINO), a fast Fourier transformation (FFT) core (FFT), and a vector core (VEC) for diverse kernels, integrated in a high-bandwidth star-ring topology. The VOTA’s frame-based instruction set and execution enable a 537 GFLOPS performance, adapt to variances of CF trackers, and reduce the code size. An instruction-chaining mechanism permits inter-core pipelining and improves the hardware utilization up to 84.2%. A 10.2-mm 2 28-nm FP16 system on chip (SoC) prototype incorporating the VOTA, an RISC-V host CPU, and other supportive peripherals is taped out and measured to achieve 2.45 TFLOPS/W at 0.72 V. Running the oriented particle CF (OPCF), a CF tracker enhanced by adaptive boosting and particle filtering, the SoC chip achieves 1157 frames/s ( $640\times480$ frame size) at 0.9 V and 175 MHz, consuming 296 mW.