Sung Kim
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
Sung Kim has a total of 43 co-authors. Here are the top co-authors by number of shared publications:
Naveen John
Department of Electrical and Computer Engineering, University of Washington, Seattle, WA, USA
3 shared publications- Venkata Rajesh Pamula
Department of Electrical and Computer Engineering, University of Washington, Seattle, WA, USA
3 shared publications Visvesh S. Sathe
Department of Electrical and Computer Engineering, University of Washington, Seattle, WA, USA
3 shared publicationsArmin Alaghi
2 shared publicationsFahim Ur Rahman
Department of Electrical and Computer Engineering, University of Washington, Seattle WA 98195 USA
2 shared publications- Fahim Ur Rahman
Department of Electrical and Computer Engineering, University of Washington, Seattle, WA 98195 USA
2 shared publications
Latest Publications
- 31 January 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 57 (4) , 986-998
Versa: A 36-Core Systolic Multiprocessor With Dynamically Reconfigurable Interconnect and Memory
We present Versa, an energy-efficient 36-core systolic multiprocessor with dynamically reconfigurable interconnects and memory. Versa leverages reconfigurable functional units and systolic-enhanced ARM cores to adapt for different algorithm characteristics, providing optimized bandwidth, access latency, and data reuse. Hardware support for crucial thread-synchronization operations enables a tree-based algorithm with 6.5$\times $ improvement in synchronization latency. Measured on a diverse set of compute kernels, Versa’s design features culminate in median energy-efficiency improvements of 37.2$\times $ and 11.6$\times $ over mobile CPU and GPU baselines, respectively.- 31 January 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 57 (4) , 986-998
Versa: A 36-Core Systolic Multiprocessor With Dynamically Reconfigurable Interconnect and Memory
We present Versa, an energy-efficient 36-core systolic multiprocessor with dynamically reconfigurable interconnects and memory. Versa leverages reconfigurable functional units and systolic-enhanced ARM cores to adapt for different algorithm characteristics, providing optimized bandwidth, access latency, and data reuse. Hardware support for crucial thread-synchronization operations enables a tree-based algorithm with 6.5$\times $ improvement in synchronization latency. Measured on a diverse set of compute kernels, Versa’s design features culminate in median energy-efficiency improvements of 37.2$\times $ and 11.6$\times $ over mobile CPU and GPU baselines, respectively.- 17 September 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 54 (11) , 3215-3225
An All-Digital Fused PLL-Buck Architecture for 82% Average V dd-Margin Reduction in a 0.6-to-1.0-V Cortex-M0 Processor
Traditional digital systems employ independent loops to control supply voltage (V dd ) and clock frequency (f clk ). A clock regulation loop, for instance, a phase-locked loop (PLL), locks the system clock to a reference clock (REFCLK). Concurrently, a voltage regulation loop sets V dd to a target value under rapidly varying load current or line-side disturbances. Limitations in voltage regulator loop bandwidth result in significant V dd fluctuation, leading to timing slack degradation in digital voltage domains. Maintaining timing slack in the presence of supply noise or temperature (T) variation incurs energy-wasteful V dd guardbands. In this article, we present a unified clock and power (UniCaP) architecture that instantaneously adapts f clk to V dd and T variations, requiring minimal guardbands. UniCaP absorbs V dd control into the clock regulation loop, using it to adjust f clk and lock system clock to REFCLK. UniCaP enables all-digital construction with aggressive V dd guardband reduction and the capability for on-the-fly dynamic voltage and frequency scaling (DVFS) events. We deployed the UniCaP architecture on a 65-nm buck converter test chip powering a 0.6-1.0-V Cortex-M0 microprocessor with autonomous transition between continuous and discontinuous conduction modes (CCM/DCM) in order to support a wide load current range. The UniCaP test chip demonstrates 82% average Vdd guardband reduction, without any performance loss from adaptive clocking, and temperature margin reductions of 40-55 mV.
- 17 September 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 54 (11) , 3215-3225
An All-Digital Fused PLL-Buck Architecture for 82% Average V dd-Margin Reduction in a 0.6-to-1.0-V Cortex-M0 Processor
Traditional digital systems employ independent loops to control supply voltage (V dd ) and clock frequency (f clk ). A clock regulation loop, for instance, a phase-locked loop (PLL), locks the system clock to a reference clock (REFCLK). Concurrently, a voltage regulation loop sets V dd to a target value under rapidly varying load current or line-side disturbances. Limitations in voltage regulator loop bandwidth result in significant V dd fluctuation, leading to timing slack degradation in digital voltage domains. Maintaining timing slack in the presence of supply noise or temperature (T) variation incurs energy-wasteful V dd guardbands. In this article, we present a unified clock and power (UniCaP) architecture that instantaneously adapts f clk to V dd and T variations, requiring minimal guardbands. UniCaP absorbs V dd control into the clock regulation loop, using it to adjust f clk and lock system clock to REFCLK. UniCaP enables all-digital construction with aggressive V dd guardband reduction and the capability for on-the-fly dynamic voltage and frequency scaling (DVFS) events. We deployed the UniCaP architecture on a 65-nm buck converter test chip powering a 0.6-1.0-V Cortex-M0 microprocessor with autonomous transition between continuous and discontinuous conduction modes (CCM/DCM) in order to support a wide load current range. The UniCaP test chip demonstrates 82% average Vdd guardband reduction, without any performance loss from adaptive clocking, and temperature margin reductions of 40-55 mV.
- 31 January 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Solid-State Circuits Letters
- Vol. 1 (12) , 237-240
A 65-nm CMOS 3.2-to-86 Mb/s 2.58 pJ/bit Highly Digital True-Random-Number Generator With Integrated De-Correlation and Bias Correction
This letter presents a highly digital, technology scalable, and energy-efficient cryptographic-quality true random number generator (TRNG). The proposed architecture presents a balanced approach to TRNG design, relying on a simpler, non-cryptographic quality physical random number generator (phyRNG) combined with energy-efficient integrated post-processing to de-correlate and de-bias the phyRNG bitstream. Operating at a supply voltage (Vdd) of 0.53 V, a 65nm CMOS prototype of the TRNG achieves a peak energy-efficiency of 2.58 pJ/bit. TRNG bitstreams pass all NIST randomness benchmarks over a Vdd range of 0.5–1.05 V across –20–100 ∘C, demonstrating its efficacy and robust operation over a wide Vdd and temperature range.
- 31 January 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Solid-State Circuits Letters
- Vol. 1 (12) , 237-240
A 65-nm CMOS 3.2-to-86 Mb/s 2.58 pJ/bit Highly Digital True-Random-Number Generator With Integrated De-Correlation and Bias Correction
This letter presents a highly digital, technology scalable, and energy-efficient cryptographic-quality true random number generator (TRNG). The proposed architecture presents a balanced approach to TRNG design, relying on a simpler, non-cryptographic quality physical random number generator (phyRNG) combined with energy-efficient integrated post-processing to de-correlate and de-bias the phyRNG bitstream. Operating at a supply voltage (Vdd) of 0.53 V, a 65nm CMOS prototype of the TRNG achieves a peak energy-efficiency of 2.58 pJ/bit. TRNG bitstreams pass all NIST randomness benchmarks over a Vdd range of 0.5–1.05 V across –20–100 ∘C, demonstrating its efficacy and robust operation over a wide Vdd and temperature range.
- 11 January 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 54 (4) , 1173-1184
A Unified Clock and Switched-Capacitor-Based Power Delivery Architecture for Variation Tolerance in Low-Voltage SoC Domains
- 11 January 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 54 (4) , 1173-1184
A Unified Clock and Switched-Capacitor-Based Power Delivery Architecture for Variation Tolerance in Low-Voltage SoC Domains
- 7 June 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 65 (12) , 4285-4298
Energy-Efficient Neural Network Acceleration in the Presence of Bit-Level Memory Errors
As a result of the increasing demand for deep neural network (DNN)-based services, efforts to develop hardware accelerators for DNNs are growing rapidly. However, while highly efficient accelerators on convolutional DNNs (ConvDNNs) have been developed, less progress has been made with regards to fully-connected DNNs. Based on analysis of bit-level SRAM errors, we propose memory adaptive training with in-situ canaries (MATIC), a methodology that enables aggressive voltage scaling of accelerator weight memories to improve the energyefficiency of DNN accelerators. To enable accurate operation with voltage overscaling, MATIC combines characteristics of SRAM bit failures with the error resilience of neural networks in a memory-adaptive training (MAT) process. Furthermore, PVT-related voltage margins are eliminated using bit-cells from synaptic weights as in-situ canaries to track runtime environmental variation. Demonstrated on a low-power DNN accelerator fabricated in 65 nm CMOS, MATIC enables up to 3.3× energy reduction versus the nominal voltage, or 18.6× application error reduction. We also perform a simulation study that extends MAT to Conv-DNNs, and characterize the accuracy impact of bit failure statistics. Finally, we develop a weight refinement algorithm to improve the performance of MAT, and show that it improves absolute accuracy by 0.8-1.3% or reduces training time by 5-10×.
- 7 June 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 65 (12) , 4285-4298
Energy-Efficient Neural Network Acceleration in the Presence of Bit-Level Memory Errors
As a result of the increasing demand for deep neural network (DNN)-based services, efforts to develop hardware accelerators for DNNs are growing rapidly. However, while highly efficient accelerators on convolutional DNNs (ConvDNNs) have been developed, less progress has been made with regards to fully-connected DNNs. Based on analysis of bit-level SRAM errors, we propose memory adaptive training with in-situ canaries (MATIC), a methodology that enables aggressive voltage scaling of accelerator weight memories to improve the energyefficiency of DNN accelerators. To enable accurate operation with voltage overscaling, MATIC combines characteristics of SRAM bit failures with the error resilience of neural networks in a memory-adaptive training (MAT) process. Furthermore, PVT-related voltage margins are eliminated using bit-cells from synaptic weights as in-situ canaries to track runtime environmental variation. Demonstrated on a low-power DNN accelerator fabricated in 65 nm CMOS, MATIC enables up to 3.3× energy reduction versus the nominal voltage, or 18.6× application error reduction. We also perform a simulation study that extends MAT to Conv-DNNs, and characterize the accuracy impact of bit failure statistics. Finally, we develop a weight refinement algorithm to improve the performance of MAT, and show that it improves absolute accuracy by 0.8-1.3% or reduces training time by 5-10×.
- 1 February 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 302-304
A combined all-digital PLL-buck slack regulation system with autonomous CCM/DCM transition control and 82% average voltage-margin reduction in a 0.6-to-1.0V cortex-M0 processor
Integrated Voltage Regulation (IVR) using buck converters enables efficient, fine-grained supply-voltage control in modern SoC domains [1]. However, existing IVR implementations face several challenges. As voltage domains continue to shrink, reduced per-domain decoupling capacitance requires rapid IVR transient response, leading to unfavorable efficiency and supply droop margin trade-offs. Additionally, digital domains exhibit a wide load current (Iload) range, requiring capabilities for autonomous transition between Continuous Conduction Mode (CCM) and Discontinuous Conduction Mode (DCM). All-digital IVR solutions are particularly desirable for ease of integration in SoCs.
- 1 February 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 302-304
A combined all-digital PLL-buck slack regulation system with autonomous CCM/DCM transition control and 82% average voltage-margin reduction in a 0.6-to-1.0V cortex-M0 processor
Integrated Voltage Regulation (IVR) using buck converters enables efficient, fine-grained supply-voltage control in modern SoC domains [1]. However, existing IVR implementations face several challenges. As voltage domains continue to shrink, reduced per-domain decoupling capacitance requires rapid IVR transient response, leading to unfavorable efficiency and supply droop margin trade-offs. Additionally, digital domains exhibit a wide load current (Iload) range, requiring capabilities for autonomous transition between Continuous Conduction Mode (CCM) and Discontinuous Conduction Mode (DCM). All-digital IVR solutions are particularly desirable for ease of integration in SoCs.
- 1 October 2017
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 177-186
Exploring computation-communication tradeoffs in camera systems
Cameras are the defacto sensor. The growing demand for real-time and low-power computer vision, coupled with trends towards high-efficiency heterogeneous systems, has given rise to a wide range of image processing acceleration techniques at the camera node and in the cloud. In this paper, we characterize two novel camera systems that use acceleration techniques to push the extremes of energy and performance scaling, and explore the computation-communication tradeoffs in their design. The first case study targets a camera system designed to detect and authenticate individual faces, running solely on energy harvested from RFID readers. We design a multi-accelerator SoC design operating in the sub-mW range, and evaluate it with real-world workloads to show performance and energy efficiency improvements over a general purpose microprocessor. The second camera system supports a 16-camera rig processing over 32 Gb/s of data to produce real-time 3D-360° virtual reality video. We design a multi-FPGA processing pipeline that outperforms CPU and GPU configurations by up to 10× in computation time, producing panoramic stereo video directly from the camera rig at 30 frames per second. We find that an early data reduction step, either before complex processing or offloading, is the most critical optimization for in-camera systems.
- 1 October 2017
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 177-186
Exploring computation-communication tradeoffs in camera systems
Cameras are the defacto sensor. The growing demand for real-time and low-power computer vision, coupled with trends towards high-efficiency heterogeneous systems, has given rise to a wide range of image processing acceleration techniques at the camera node and in the cloud. In this paper, we characterize two novel camera systems that use acceleration techniques to push the extremes of energy and performance scaling, and explore the computation-communication tradeoffs in their design. The first case study targets a camera system designed to detect and authenticate individual faces, running solely on energy harvested from RFID readers. We design a multi-accelerator SoC design operating in the sub-mW range, and evaluate it with real-world workloads to show performance and energy efficiency improvements over a general purpose microprocessor. The second camera system supports a 16-camera rig processing over 32 Gb/s of data to produce real-time 3D-360° virtual reality video. We design a multi-FPGA processing pipeline that outperforms CPU and GPU configurations by up to 10× in computation time, producing panoramic stereo video directly from the camera rig at 30 frames per second. We find that an early data reduction step, either before complex processing or offloading, is the most critical optimization for in-camera systems.
- 17 August 2016
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 18 (4) , 890-901
Real-Time Bidirectional Traffic Flow Parameter Estimation From Aerial Videos
Unmanned aerial vehicles (UAVs) are gaining popularity in traffic monitoring due to their low cost, high flexibility, and wide view range. Traffic flow parameters such as speed, density, and volume extracted from UAV-based traffic videos are critical for traffic state estimation and traffic control and have recently received much attention from researchers. However, different from stationary surveillance videos, the camera platforms move with UAVs, and the background motion in aerial videos makes it very challenging to process for data extraction. To address this problem, a novel framework for real-time traffic flow parameter estimation from aerial videos is proposed. The proposed system identifies the directions of traffic streams and extracts traffic flow parameters of each traffic stream separately. Our method incorporates four steps that make use of the Kanade-Lucas-Tomasi (KLT) tracker, k-means clustering, connected graphs, and traffic flow theory. The KLT tracker and k-means clustering are used for interest-point-based motion analysis; then, four constraints are proposed to further determine the connectivity of interest points belonging to one traffic stream cluster. Finally, the average speed of a traffic stream as well as density and volume can be estimated using outputs from previous steps and reference markings. Our method was tested on five videos taken in very different scenarios. The experimental results show that in our case studies, the proposed method achieves about 96% and 87% accuracy in estimating average traffic stream speed and vehicle count, respectively. The method also achieves a fast processing speed that enables real-time traffic information estimation.
- 17 August 2016
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 18 (4) , 890-901
Real-Time Bidirectional Traffic Flow Parameter Estimation From Aerial Videos
Unmanned aerial vehicles (UAVs) are gaining popularity in traffic monitoring due to their low cost, high flexibility, and wide view range. Traffic flow parameters such as speed, density, and volume extracted from UAV-based traffic videos are critical for traffic state estimation and traffic control and have recently received much attention from researchers. However, different from stationary surveillance videos, the camera platforms move with UAVs, and the background motion in aerial videos makes it very challenging to process for data extraction. To address this problem, a novel framework for real-time traffic flow parameter estimation from aerial videos is proposed. The proposed system identifies the directions of traffic streams and extracts traffic flow parameters of each traffic stream separately. Our method incorporates four steps that make use of the Kanade-Lucas-Tomasi (KLT) tracker, k-means clustering, connected graphs, and traffic flow theory. The KLT tracker and k-means clustering are used for interest-point-based motion analysis; then, four constraints are proposed to further determine the connectivity of interest points belonging to one traffic stream cluster. Finally, the average speed of a traffic stream as well as density and volume can be estimated using outputs from previous steps and reference markings. Our method was tested on five videos taken in very different scenarios. The experimental results show that in our case studies, the proposed method achieves about 96% and 87% accuracy in estimating average traffic stream speed and vehicle count, respectively. The method also achieves a fast processing speed that enables real-time traffic information estimation.
- 1 October 2015
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-5
Motion-vector clustering for traffic speed detection from UAV video
A novel method for detecting the average speed of traffic from non-stationary aerial video is presented. The method first extracts interest points from a pair of frames and performs interest point tracking with an optical flow algorithm. The output of the optical flow is a set of motion vectors which are k-means clustered in velocity space. The centers of the clusters correspond to the average velocities of traffic and the background, and are used to determine the speed of traffic relative to the background. The proposed method is tested on a 70-frame test sequence of UAV aerial video, and achieves an average error for speed estimates of less than 12%.
- 1 October 2015
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-5
Motion-vector clustering for traffic speed detection from UAV video
A novel method for detecting the average speed of traffic from non-stationary aerial video is presented. The method first extracts interest points from a pair of frames and performs interest point tracking with an optical flow algorithm. The output of the optical flow is a set of motion vectors which are k-means clustered in velocity space. The centers of the clusters correspond to the average velocities of traffic and the background, and are used to determine the speed of traffic relative to the background. The proposed method is tested on a 70-frame test sequence of UAV aerial video, and achieves an average error for speed estimates of less than 12%.