Johnny Rhe
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Johnny Rhe has a total of 13 co-authors. Here are the top co-authors by number of shared publications:
- Jong Hwan Ko
Sungkyunkwan University
5 shared publications - Donguk Seo
Department of Electrical and Computer Engineering, Sungkyunkwan University, Suwon, South Korea
2 shared publications - Eunyoung Lee
Sungkyunkwan University, Suwon, Korea
2 shared publications Gicheol Shin
Sungkyunkwan University,Dept. of Electrical and Computer Engineering,Suwon,Korea
2 shared publications- Jaerok Kim
Department of Electrical and Computer Engineering, Sungkyunkwan University, Suwon, South Korea
2 shared publications - Seonho Kim
Samsung Electronics
2 shared publications
Latest Publications
- 22 February 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 71 (12) , 6138-6151
KERNTROL: Kernel Shape Control Toward Ultimate Memory Utilization for In-Memory Convolutional Weight Mapping
Processing-in-memory (PIM) architectures have been highlighted as one of the most viable options for faster and more power-efficient computation. Paired with a convolutional weight mapping scheme, PIM arrays can accelerate various deep convolutional neural networks (CNNs) and the applications that adopt them. Recently, shift and duplicate kernel (SDK) convolutional weight mapping scheme was proposed, achieving up to 50% throughput improvement over the prior arts. However, the traditional pattern-based pruning methods, which were adopted for row-skipping and computing cycle reduction, are not optimal for the latest SDK mapping due to the loss of structural regularity caused by the shifted and duplicated kernels. To address this challenge, we propose kernel shape control (KERNTROL), a method where kernel shapes are controlled depending on their mapped columns with the purpose of fostering a structural regularity that is favorable in achieving a high row-skipping ratio and model accuracy. Instead of permanently pruning the weights, KERNTROL with an empty mask (KERNTROL-M) temporarily omits them in the underutilized row using a utilization threshold, thereby preserving important weight elements. However, a significant portion of the memory cells is still underutilized where the threshold is not enforced. To overcome this, we extend KERNTROL-M into KERNTROL with compensatory weights (KERNTROL-C). By populating idle cells with compensatory weights, KERNTROL-C can offset the accuracy drop from weight omission. In comparison to pattern-based pruning approaches, KERNTROL-C achieves simultaneous improvements of up to 36.4% improvement in the compression rate and 5% in model accuracy with up to 100% array utilization.
- 22 February 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 71 (12) , 6138-6151
KERNTROL: Kernel Shape Control Toward Ultimate Memory Utilization for In-Memory Convolutional Weight Mapping
Processing-in-memory (PIM) architectures have been highlighted as one of the most viable options for faster and more power-efficient computation. Paired with a convolutional weight mapping scheme, PIM arrays can accelerate various deep convolutional neural networks (CNNs) and the applications that adopt them. Recently, shift and duplicate kernel (SDK) convolutional weight mapping scheme was proposed, achieving up to 50% throughput improvement over the prior arts. However, the traditional pattern-based pruning methods, which were adopted for row-skipping and computing cycle reduction, are not optimal for the latest SDK mapping due to the loss of structural regularity caused by the shifted and duplicated kernels. To address this challenge, we propose kernel shape control (KERNTROL), a method where kernel shapes are controlled depending on their mapped columns with the purpose of fostering a structural regularity that is favorable in achieving a high row-skipping ratio and model accuracy. Instead of permanently pruning the weights, KERNTROL with an empty mask (KERNTROL-M) temporarily omits them in the underutilized row using a utilization threshold, thereby preserving important weight elements. However, a significant portion of the memory cells is still underutilized where the threshold is not enforced. To overcome this, we extend KERNTROL-M into KERNTROL with compensatory weights (KERNTROL-C). By populating idle cells with compensatory weights, KERNTROL-C can offset the accuracy drop from weight omission. In comparison to pattern-based pruning approaches, KERNTROL-C achieves simultaneous improvements of up to 36.4% improvement in the compression rate and 5% in model accuracy with up to 100% array utilization.
- 9 May 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal on Emerging and Selected Topics in Circuits and Systems
- Vol. 12 (2) , 408-421
VWC-SDK: Convolutional Weight Mapping Using Shifted and Duplicated Kernel With Variable Windows and Channels
With their high energy efficiency, processing-in-memory (PIM) arrays are being increasingly used for the convolutional neural network (CNN) inference. In the PIM-based CNN inference, the computational latency and energy are dependent on how the CNN weights are mapped to the PIM array. A recent study proposed the shifted and duplicated kernel (SDK) mapping that reuses input feature maps with a unit of a parallel window, which is convolved with duplicated kernels to obtain multiple output elements in parallel. However, the existing SDK-based mapping algorithm does not always result in minimum computing cycles because it only maps a square-shaped parallel window with the entire channels. In this paper, we introduce a novel mapping algorithm called variable-window SDK (VW-SDK), which adaptively determines the shape of the parallel window that leads to the minimum computing cycles for the given convolutional layer and PIM array. By allowing rectangular-shaped windows with partial channels, VW-SDK utilizes the PIM array more efficiently, thereby further reducing the number of computing cycles. To further remove the inefficient computing cycles caused by the residual channels, we extend the VW-SDK algorithm into VWC-SDK (SDK with variable windows and channels) that additionally performs residual channel pruning. The simulation with a$512\times 512$ PIM array and ResNet-20 shows that VW-SDK improves the inference speed by$1.29\times $ compared to the existing SDK-based algorithm. The results also show that residual channel pruning improves the inference speed of ResNet-20 by up to$\sim 1.38\times $ when compared to the original network without pruning.- 9 May 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal on Emerging and Selected Topics in Circuits and Systems
- Vol. 12 (2) , 408-421
VWC-SDK: Convolutional Weight Mapping Using Shifted and Duplicated Kernel With Variable Windows and Channels
With their high energy efficiency, processing-in-memory (PIM) arrays are being increasingly used for the convolutional neural network (CNN) inference. In the PIM-based CNN inference, the computational latency and energy are dependent on how the CNN weights are mapped to the PIM array. A recent study proposed the shifted and duplicated kernel (SDK) mapping that reuses input feature maps with a unit of a parallel window, which is convolved with duplicated kernels to obtain multiple output elements in parallel. However, the existing SDK-based mapping algorithm does not always result in minimum computing cycles because it only maps a square-shaped parallel window with the entire channels. In this paper, we introduce a novel mapping algorithm called variable-window SDK (VW-SDK), which adaptively determines the shape of the parallel window that leads to the minimum computing cycles for the given convolutional layer and PIM array. By allowing rectangular-shaped windows with partial channels, VW-SDK utilizes the PIM array more efficiently, thereby further reducing the number of computing cycles. To further remove the inefficient computing cycles caused by the residual channels, we extend the VW-SDK algorithm into VWC-SDK (SDK with variable windows and channels) that additionally performs residual channel pruning. The simulation with a$512\times 512$ PIM array and ResNet-20 shows that VW-SDK improves the inference speed by$1.29\times $ compared to the existing SDK-based algorithm. The results also show that residual channel pruning improves the inference speed of ResNet-20 by up to$\sim 1.38\times $ when compared to the original network without pruning.- 13 September 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 123-126
A Charge-Domain Computation-In-Memory Macro with Versatile All-Around-Wire-Capacitor for Variable-Precision Computation and Array-Embedded DA/AD Conversions
A charge-domain SRAM-based CIM macro is proposed to implement efficient neural accelerator with variable input/weight/output precisions. A versatile All-Around-Wire-Capacitor (AAWC) is introduced to enable accurate charge-domain MAV operation with 1–5 bit variable-precision weight capability with 10T SRAM cell structure. The AAWC also allows variable-precision array-embedded 1–5 bit digital-to-analog (DA) conversion for input, and 1–6 bit analog-to-digital (AD) conversion for output without additional passive devices. Such array-embedded DA/AD conversion scheme significantly reduces area overhead and simplifies peripheral circuits to achieve high array efficiency while achieving high energy efficiency of 24–330 TOPS/W.
- 13 September 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 123-126
A Charge-Domain Computation-In-Memory Macro with Versatile All-Around-Wire-Capacitor for Variable-Precision Computation and Array-Embedded DA/AD Conversions
A charge-domain SRAM-based CIM macro is proposed to implement efficient neural accelerator with variable input/weight/output precisions. A versatile All-Around-Wire-Capacitor (AAWC) is introduced to enable accurate charge-domain MAV operation with 1–5 bit variable-precision weight capability with 10T SRAM cell structure. The AAWC also allows variable-precision array-embedded 1–5 bit digital-to-analog (DA) conversion for input, and 1–6 bit analog-to-digital (AD) conversion for output without additional passive devices. Such array-embedded DA/AD conversion scheme significantly reduces area overhead and simplifies peripheral circuits to achieve high array efficiency while achieving high energy efficiency of 24–330 TOPS/W.
- 13 September 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 123-126
A Charge-Domain Computation-In-Memory Macro with Versatile All-Around-Wire-Capacitor for Variable-Precision Computation and Array-Embedded DA/AD Conversions
A charge-domain SRAM-based CIM macro is proposed to implement efficient neural accelerator with variable input/weight/output precisions. A versatile All-Around-Wire-Capacitor (AAWC) is introduced to enable accurate charge-domain MAV operation with 1-5 bit variable-precision weight capability with 10T SRAM cell structure. The AAWC also allows variable-precision array-embedded 1-5 bit digital-to-analog (DA) conversion for input, and 1-6 bit analog-to-digital (AD) conversion for output without additional passive devices. Such array-embedded DA/AD conversion scheme significantly reduces area overhead and simplifies peripheral circuits to achieve high array efficiency while achieving high energy efficiency of 24-330 TOPS/W.
- 13 September 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 123-126
A Charge-Domain Computation-In-Memory Macro with Versatile All-Around-Wire-Capacitor for Variable-Precision Computation and Array-Embedded DA/AD Conversions
A charge-domain SRAM-based CIM macro is proposed to implement efficient neural accelerator with variable input/weight/output precisions. A versatile All-Around-Wire-Capacitor (AAWC) is introduced to enable accurate charge-domain MAV operation with 1-5 bit variable-precision weight capability with 10T SRAM cell structure. The AAWC also allows variable-precision array-embedded 1-5 bit digital-to-analog (DA) conversion for input, and 1-6 bit analog-to-digital (AD) conversion for output without additional passive devices. Such array-embedded DA/AD conversion scheme significantly reduces area overhead and simplifies peripheral circuits to achieve high array efficiency while achieving high energy efficiency of 24-330 TOPS/W.