Atsutake Kosuge
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
Atsutake Kosuge has a total of 54 co-authors. Here are the top co-authors by number of shared publications:
- Mototsugu Hamada
Graduate School of Engineering, The University of Tokyo, Tokyo, Japan
16 shared publications - Tadahiro Kuroda
Graduate School of Engineering, The University of Tokyo, Tokyo, Japan
14 shared publications - Kota Shiba
Taiwan Semiconductor Manufacturing Company, Yokohama, Kanagawa, Japan
7 shared publications Mitsuji Okada
Graduate School of Engineering, The University of Tokyo, Tokyo, Japan
3 shared publicationsYao-Chung Hsu
Graduate School of Engineering, University of Tokyo, Bunkyo, Japan
3 shared publications- Jaewon Shin
Graduate School of Engineering, University of Tokyo, Tokyo, Japan
2 shared publications
Latest Publications
- Open Access
- 2 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-3
Designing Chips in the Browser: Democratizing ASIC Prototyping with Gate Arrays
- 2 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-3
Open AccessDesigning Chips in the Browser: Democratizing ASIC Prototyping with Gate Arrays
- 16 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. PP (15498328) , 1-11
A 28-nm 0.8M-Weights/mm 2 9.1-TOPS/mm 2 All-Analog SRAM-Based Compute-in-Memory Macro Using Fine-Grained Structured Pruning With Adaptive-Ranging ADC
A 0.8M-weights/mm ${}^{2}~9.1$ -TOPS/mm2 SRAM-based all-analog CIM macro for AI edge devices has been developed in 28-nm CMOS. This paper presents the CIM for both efficient storing and processing of sparse weight matrices using fine-grained structured pruning which achieves a $5.5\times $ improvement in both memory and compute densities. All-analog operation from MAC, biasing, ReLU to layer-to-layer quantization eliminates digital post-processing circuits and achieves high system-level energy efficiency and low latency. It is enabled by a PVT-tracking adaptive-ranging ADC with sensing margin improved by up to $576\times $ . The proposed CIM achieves $\gt 6\times $ higher normalized memory and compute density than other CIMs with 116-TOPS/W system-level energy efficiency and low latency.
- 16 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. PP (15498328) , 1-11
A 28-nm 0.8M-Weights/mm 2 9.1-TOPS/mm 2 All-Analog SRAM-Based Compute-in-Memory Macro Using Fine-Grained Structured Pruning With Adaptive-Ranging ADC
A 0.8M-weights/mm ${}^{2}~9.1$ -TOPS/mm2 SRAM-based all-analog CIM macro for AI edge devices has been developed in 28-nm CMOS. This paper presents the CIM for both efficient storing and processing of sparse weight matrices using fine-grained structured pruning which achieves a $5.5\times $ improvement in both memory and compute densities. All-analog operation from MAC, biasing, ReLU to layer-to-layer quantization eliminates digital post-processing circuits and achieves high system-level energy efficiency and low latency. It is enabled by a PVT-tracking adaptive-ranging ADC with sensing margin improved by up to $576\times $ . The proposed CIM achieves $\gt 6\times $ higher normalized memory and compute density than other CIMs with 116-TOPS/W system-level energy efficiency and low latency.
- Open Access
- 23 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of Circuits and Systems
- Vol. 7, 11-20
A Connectivity-Aware Via-Programmable DNN Processor Using a Single Photomask
This paper presents a via-programmable DNN processor architecture, the Via-Programmable Neuron Array (VPNA), designed for low-NRE and low-power AIoT applications. To enable shared base-chip layouts across diverse workloads, a connectivity-aware design ensures tile-to-tile routing under a column-wise placement rule. A 6× 6 programmable-wire structure supports task-specific data paths, and via-based ternary-weight mapping allows multiple tasks to reuse the same base chip with a single via mask. A unified bit-serial neuron circuit supports convolution and pooling operations under both neuron-serial and neuron-parallel modes, completing the functional implementation required for one-dimensional time-series DNNs. Post-layout evaluations in a 40 nm CMOS process demonstrate sub-milliwatt power consumption and sufficient inference accuracy across representative AIoT tasks, including keyword spotting, ECG arrhythmia detection, and EEG seizure detection. Compared with prior FPGA-and ASIC-based accelerators, the proposed architecture achieves a better trade-off among low power, low NRE cost, and task-level flexibility, highlighting its potential as a scalable foundation for future ultra-low-NRE and field-programmable AIoT processors.
- 23 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of Circuits and Systems
- Vol. 7, 11-20
Open AccessA Connectivity-Aware Via-Programmable DNN Processor Using a Single Photomask
This paper presents a via-programmable DNN processor architecture, the Via-Programmable Neuron Array (VPNA), designed for low-NRE and low-power AIoT applications. To enable shared base-chip layouts across diverse workloads, a connectivity-aware design ensures tile-to-tile routing under a column-wise placement rule. A 6× 6 programmable-wire structure supports task-specific data paths, and via-based ternary-weight mapping allows multiple tasks to reuse the same base chip with a single via mask. A unified bit-serial neuron circuit supports convolution and pooling operations under both neuron-serial and neuron-parallel modes, completing the functional implementation required for one-dimensional time-series DNNs. Post-layout evaluations in a 40 nm CMOS process demonstrate sub-milliwatt power consumption and sufficient inference accuracy across representative AIoT tasks, including keyword spotting, ECG arrhythmia detection, and EEG seizure detection. Compared with prior FPGA-and ASIC-based accelerators, the proposed architecture achieves a better trade-off among low power, low NRE cost, and task-level flexibility, highlighting its potential as a scalable foundation for future ultra-low-NRE and field-programmable AIoT processors.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of the Solid-State Circuits Society
- Vol. 6, 439-449
A Wired-Logic and Near-Memory Hybrid Processor for General-Purpose Audio Inference Using a Self-Supervised Encoder
This work presents a wired-logic (WL) and near-memory (NM) hybrid processor that enables general-purpose audio inference by embedding a compact self-supervised learning (SSL) pretrained encoder into an area-efficient WL core while retaining task-level adaptability through an NM-based classifier. To obtain a compact encoder suitable for WL implementation, we develop an NNN-SSL framework that trains a nonlinear neural network (NNN) using ternary weights, 80% pruning, and 14-bit activations within an augmentation-based SSL pipeline. The resulting 64-channel encoder provides unified feature extraction across speech, environmental sound, and musical-instrument domains. A global-pooling block and a frequency-serial fully connected (FC) architecture support variable-length audio inputs and reduce the FC area from 3.56mm2 to 0.76mm2. A 40 nm test chip achieves 48.1 TOPS/W for the convolution layer, while the full-network energy is estimated at 0.87 μJ/inference on GSCD-12. Furthermore, scaling the SSL pretraining dataset and the number of epochs enables accuracy improvements under a fixed-area constraint, showing that SSL can compensate for limited model capacity in WL-based designs. By combining SSL-driven general-purpose feature learning with WL-centric energy efficiency and minimal NM programmability, the proposed processor demonstrates a practical path toward compact, low-power, multi-domain audio AI for edge devices.
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of the Solid-State Circuits Society
- Vol. 6, 439-449
Open AccessA Wired-Logic and Near-Memory Hybrid Processor for General-Purpose Audio Inference Using a Self-Supervised Encoder
This work presents a wired-logic (WL) and near-memory (NM) hybrid processor that enables general-purpose audio inference by embedding a compact self-supervised learning (SSL) pretrained encoder into an area-efficient WL core while retaining task-level adaptability through an NM-based classifier. To obtain a compact encoder suitable for WL implementation, we develop an NNN-SSL framework that trains a nonlinear neural network (NNN) using ternary weights, 80% pruning, and 14-bit activations within an augmentation-based SSL pipeline. The resulting 64-channel encoder provides unified feature extraction across speech, environmental sound, and musical-instrument domains. A global-pooling block and a frequency-serial fully connected (FC) architecture support variable-length audio inputs and reduce the FC area from 3.56mm2 to 0.76mm2. A 40 nm test chip achieves 48.1 TOPS/W for the convolution layer, while the full-network energy is estimated at 0.87 μJ/inference on GSCD-12. Furthermore, scaling the SSL pretraining dataset and the number of epochs enables accuracy improvements under a fixed-area constraint, showing that SSL can compensate for limited model capacity in WL-based designs. By combining SSL-driven general-purpose feature learning with WL-centric energy efficiency and minimal NM programmability, the proposed processor demonstrates a practical path toward compact, low-power, multi-domain audio AI for edge devices.
- 25 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Components, Packaging and Manufacturing Technology
- Vol. 16 (5) , 991-1000
A Thermal and System Analysis of SoC-Memory 3-D Stacking Using Massive Orthogonal Stacking Assembly
A novel chip stacking method with low thermal resistance for 3D integration of a large number of memory chips and processor chips is proposed, namely Massive Orthogonal Stacking Assembly of IC (MOSAIC) technology. Through thermal resistance analysis, it is possible to achieve a high effective thermal conductivity of 0.88 W/(cm·K) and low operational temperature with 60 chips stacked on a GPU. This represents three times the number of chips compared to conventional parallel chip stacking. While MOSAIC exhibits good heat dissipation capability, there is a strong trade-off between heat dissipation capability and system performance. To gain a systematic understanding of the MOSAIC system, a thermal model analysis is developed. This provides designers with an optimal solution for MOSAIC design that satisfies both good heat dissipation capability and SoC performance.
- 25 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Components, Packaging and Manufacturing Technology
- Vol. 16 (5) , 991-1000
A Thermal and System Analysis of SoC-Memory 3-D Stacking Using Massive Orthogonal Stacking Assembly
A novel chip stacking method with low thermal resistance for 3D integration of a large number of memory chips and processor chips is proposed, namely Massive Orthogonal Stacking Assembly of IC (MOSAIC) technology. Through thermal resistance analysis, it is possible to achieve a high effective thermal conductivity of 0.88 W/(cm·K) and low operational temperature with 60 chips stacked on a GPU. This represents three times the number of chips compared to conventional parallel chip stacking. While MOSAIC exhibits good heat dissipation capability, there is a strong trade-off between heat dissipation capability and system performance. To gain a systematic understanding of the MOSAIC system, a thermal model analysis is developed. This provides designers with an optimal solution for MOSAIC design that satisfies both good heat dissipation capability and SoC performance.
- Open Access
- 20 January 2025
- Published by Association for Computing Machinery (ACM)
- p. 325-327
A Coarse- and Fine-Grained LUT Segmentation Method Enabling Single FPGA Implementation of Wired-Logic DNN Processor
A coarse- and fine-grained LUT segmentation technique is developed for wired-logic AI processors to improve FPGA resource utilization efficiency. By applying the proposed technique to FPGA-based wired-logic processors used for CIFAR-10 classification and keyword spotting, the hardware resource requirements for nonlinear functions were reduced by 92% and 92.8%, respectively, with negligible accuracy degradation.
- 20 January 2025
- Published by Association for Computing Machinery (ACM)
- p. 325-327
Open AccessA Coarse- and Fine-Grained LUT Segmentation Method Enabling Single FPGA Implementation of Wired-Logic DNN Processor
A coarse- and fine-grained LUT segmentation technique is developed for wired-logic AI processors to improve FPGA resource utilization efficiency. By applying the proposed technique to FPGA-based wired-logic processors used for CIFAR-10 classification and keyword spotting, the hardware resource requirements for nonlinear functions were reduced by 92% and 92.8%, respectively, with negligible accuracy degradation.
- 13 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Solid-State Circuits Magazine
- Vol. 16 (4) , 111-113
Solid-State Circuit Directions Committee’s “Think Impact With ICs” Workshop on System, Circuit, Device, and Packaging Co-Optimization for Next Generation AI Systems [Society News]
Provides society information that may include news, reviews or technical notes that should be of interest to practitioners and researchers.
- 13 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Solid-State Circuits Magazine
- Vol. 16 (4) , 111-113
Solid-State Circuit Directions Committee’s “Think Impact With ICs” Workshop on System, Circuit, Device, and Packaging Co-Optimization for Next Generation AI Systems [Society News]
Provides society information that may include news, reviews or technical notes that should be of interest to practitioners and researchers.
- 6 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 33 (3) , 746-756
Agile-X: A Structured-ASIC Created With a Mask-Less Lithography System Enabling Low-Cost and Agile Chip Fabrication
Scaling to finer CMOS process nodes necessitates more masks, resulting in higher costs and extended turnaround times (TATs). High costs and long TATs have hindered researchers outside the field of integrated circuits, including those in medicine, physics, and science from prototyping their own chips. Therefore, opportunities for diverse innovations in integrated circuits and talent development have been limited. We have developed the Agile-X platform for low-cost, rapid manufacturing of system-on-chips. Users can implement their own dedicated circuits with gate-array circuits on a base chip, which has common intellectual properties (IPs) such as RISC-V CPUs, various IOs, and ADCs. The base chip is manufactured in a foundry up to the intermediate metal layers and shipped with metal deposition on its surface. By directly drawing wiring patterns on this base chip with a mask-less lithography system, custom chips can be manufactured on-site without masks. As this process only requires wiring and eliminates masks, production time is drastically reduced compared to traditional full-mask wafer processes and multiproject wafer (MPW) shuttles. Development and manufacturing costs for the base chip, including preintegrated IPs, are shared among all Agile-X users. This reduces both IP and base-chip wafer costs per user. We prototyped wafers using a 0.18- $\mu $ m CMOS process and tested the proposed structured ASIC platform and manufacturing process using mask-less lithography systems. The results indicate that the process from inputting GDS data to lithography and dry etching can be completed within 30 min, and custom application-specific integrated circuits (ASICs) can be manufactured within a day. Compared with full-mask wafer design and manufacturing, the manufacturing cost per chip, including IP costs, is reduced from 271000 USD to 22 USD, a reduction of 1/12252, and the manufacturing period is reduced from 20 days to 30 min, a reduction of 1/960.
- 6 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 33 (3) , 746-756
Agile-X: A Structured-ASIC Created With a Mask-Less Lithography System Enabling Low-Cost and Agile Chip Fabrication
Scaling to finer CMOS process nodes necessitates more masks, resulting in higher costs and extended turnaround times (TATs). High costs and long TATs have hindered researchers outside the field of integrated circuits, including those in medicine, physics, and science from prototyping their own chips. Therefore, opportunities for diverse innovations in integrated circuits and talent development have been limited. We have developed the Agile-X platform for low-cost, rapid manufacturing of system-on-chips. Users can implement their own dedicated circuits with gate-array circuits on a base chip, which has common intellectual properties (IPs) such as RISC-V CPUs, various IOs, and ADCs. The base chip is manufactured in a foundry up to the intermediate metal layers and shipped with metal deposition on its surface. By directly drawing wiring patterns on this base chip with a mask-less lithography system, custom chips can be manufactured on-site without masks. As this process only requires wiring and eliminates masks, production time is drastically reduced compared to traditional full-mask wafer processes and multiproject wafer (MPW) shuttles. Development and manufacturing costs for the base chip, including preintegrated IPs, are shared among all Agile-X users. This reduces both IP and base-chip wafer costs per user. We prototyped wafers using a 0.18- $\mu $ m CMOS process and tested the proposed structured ASIC platform and manufacturing process using mask-less lithography systems. The results indicate that the process from inputting GDS data to lithography and dry etching can be completed within 30 min, and custom application-specific integrated circuits (ASICs) can be manufactured within a day. Compared with full-mask wafer design and manufacturing, the manufacturing cost per chip, including IP costs, is reduced from 271000 USD to 22 USD, a reduction of 1/12252, and the manufacturing period is reduced from 20 days to 30 min, a reduction of 1/960.
- Open Access
- 7 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of the Electron Devices Society
- Vol. 12, 594-604
A 1.1-nJ/Classification True Analog Current Computing on Multilayer Neural Network With Crystalline-IGZO/Si-CMOS Monolithic Stack Technology
We prototyped a true analog current computing multilayer neural network (NN) chip, where multiple analog in-memory computing (AiMC) circuit blocks are connected to each other via simple analog non-linear operation circuits. The true analog current computing is achieved with the invention of an analog current rectified linear unit (ReLU) circuit of a three-stage current mirror. With the prototyped NN chip, we demonstrated that the true analog computing (1) achieves process variation compensation utilizing current driving, (2) eliminates digital-analog or analog-digital data conversion between NNs, and (3) realizes low power inference, not only in multiply-accumulate (MAC) but in ReLU operation. Through classification of Mixed National Institute of Standards and Technology dataset, the chip exhibits a low energy of 1.1 nJ/classification and an accuracy of 91.6%, achieves weight retention of five hours, much longer than dynamic random access memory, and enables 68% power reduction compared with serially connected two single-layer NN chips with analog-digital converters and digital-analog converters in between. Although periodic refresh from an external storage class memory is necessary for applications that require continuous operation exceeding five hours, our AiMC capable of MAC and non-linear operations with low power is effective in applications such as edge artificial intelligence terminals with limited power sources.
- 7 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of the Electron Devices Society
- Vol. 12, 594-604
Open AccessA 1.1-nJ/Classification True Analog Current Computing on Multilayer Neural Network With Crystalline-IGZO/Si-CMOS Monolithic Stack Technology
We prototyped a true analog current computing multilayer neural network (NN) chip, where multiple analog in-memory computing (AiMC) circuit blocks are connected to each other via simple analog non-linear operation circuits. The true analog current computing is achieved with the invention of an analog current rectified linear unit (ReLU) circuit of a three-stage current mirror. With the prototyped NN chip, we demonstrated that the true analog computing (1) achieves process variation compensation utilizing current driving, (2) eliminates digital-analog or analog-digital data conversion between NNs, and (3) realizes low power inference, not only in multiply-accumulate (MAC) but in ReLU operation. Through classification of Mixed National Institute of Standards and Technology dataset, the chip exhibits a low energy of 1.1 nJ/classification and an accuracy of 91.6%, achieves weight retention of five hours, much longer than dynamic random access memory, and enables 68% power reduction compared with serially connected two single-layer NN chips with analog-digital converters and digital-analog converters in between. Although periodic refresh from an external storage class memory is necessary for applications that require continuous operation exceeding five hours, our AiMC capable of MAC and non-linear operations with low power is effective in applications such as edge artificial intelligence terminals with limited power sources.
- 21 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Solid-State Circuits Letters
- Vol. 7, 22-25
A183.4-nJ/Inference 152.8-μW 35-Voice Commands Recognition Wired-Logic Processor Using Algorithm-Circuit Co-Optimization Technique
A 183.4-nJ/inference single-chip wired-logic DNN processor that is capable of recognizing all 35 commands defined in the industrial standard voice recognition data set (Google Speech Command Dataset) is developed. The algorithm-circuit cooptimized processor recognizes 3.5 times more commands with 1.6 times better energy efficiency than the state-of-the-art analog processor while keeping design cost low. By implementing all the processing circuits and wiring required for the 16-layer DNN onto a single chip (7.63 mm2 in 40 nm), the need to store weight coefficients and intermediate data in DRAM/SRAM is eliminated. Owing to the proposed architecture, a low power consumption of 152.8 μW is achieved, which is low enough for always-on applications on battery-powered IoT devices.
- 21 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Solid-State Circuits Letters
- Vol. 7, 22-25
A183.4-nJ/Inference 152.8-μW 35-Voice Commands Recognition Wired-Logic Processor Using Algorithm-Circuit Co-Optimization Technique
A 183.4-nJ/inference single-chip wired-logic DNN processor that is capable of recognizing all 35 commands defined in the industrial standard voice recognition data set (Google Speech Command Dataset) is developed. The algorithm-circuit cooptimized processor recognizes 3.5 times more commands with 1.6 times better energy efficiency than the state-of-the-art analog processor while keeping design cost low. By implementing all the processing circuits and wiring required for the 16-layer DNN onto a single chip (7.63 mm2 in 40 nm), the need to store weight coefficients and intermediate data in DRAM/SRAM is eliminated. Owing to the proposed architecture, a low power consumption of 152.8 μW is achieved, which is low enough for always-on applications on battery-powered IoT devices.