Michihiro Koibuchi
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
Michihiro Koibuchi has a total of 60 co-authors. Here are the top co-authors by number of shared publications:
Hideharu Amano
Keio University
5 shared publications- Hiroki Matsutani
Keio University
4 shared publications - Hideharu Amano
Department of Open Environment Science, Graduate School of Science and Technology, Keio University, Tokyo 223-8522, Japan
3 shared publications - Ikki Fujiwara
National Institute of Informatics, Japan
3 shared publications - Kenji Mizutani
National Institute of Advanced Industrial Science and Technology, Japan
3 shared publications - Naoto Fukumoto
Fujitsu Limited, Nakahara-ku, Japan
3 shared publications
Latest Publications
- Open Access
- 17 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 111-117
Tuning Optical Modulation at Floating-Point Bit Granularity for Approximate MPI Communication
High-bandwidth optical interconnects increasingly use multilevel modulation, but reduced noise margins often require Forward Error Correction (FEC), whose latency can hurt latency-sensitive MPI applications. We propose bit-position-aware optical modulation tuning that assigns different modulation configurations to floating-point bit groups according to their numerical importance. An offline auto-tuner searches modulation policies and rejects configurations that fail application-level validation. SimGrid-based MPI simulations on three NAS Parallel Benchmarks show a 1.44× improvement in expected useful performance over a conventional reliable network, where failed executions contribute zero useful performance.
- 17 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 111-117
Open AccessTuning Optical Modulation at Floating-Point Bit Granularity for Approximate MPI Communication
High-bandwidth optical interconnects increasingly use multilevel modulation, but reduced noise margins often require Forward Error Correction (FEC), whose latency can hurt latency-sensitive MPI applications. We propose bit-position-aware optical modulation tuning that assigns different modulation configurations to floating-point bit groups according to their numerical importance. An offline auto-tuner searches modulation policies and rejects configurations that fail application-level validation. SimGrid-based MPI simulations on three NAS Parallel Benchmarks show a 1.44× improvement in expected useful performance over a conventional reliable network, where failed executions contribute zero useful performance.
- Open Access
- 10 August 2026
- Published by Wiley in Concurrency and Computation: Practice and Experience
- Vol. 38 (16) , e70895
Reproducible Ozaki‐Based Matrix Multiplication on Photonic Matrix Processors With Algorithm‐Based Fault Tolerance
Photonic matrix processors are attracting attention in applications such as AI inference due to their low‐power and low‐latency properties. However, their applicability has been largely limited to low‐precision computation due to inherent stochastic analog noise and device non‐idealities. This limitation creates a large gap between the capabilities of photonic matrix processors and the potential demand for high‐precision computation in scientific computing. In this study, we propose a method to reproduce the results of the Ozaki‐based high‐precision matrix multiplication on low‐precision photonic matrix processors by combining the Ozaki scheme with a simple Algorithm‐Based Fault Tolerance (ABFT). The Ozaki scheme decomposes high‐precision operations into multiple low‐precision computations, while the ABFT introduces algorithm‐level redundancy to detect and correct errors caused by stochastic analog noise. Evaluation results show that ABFT significantly reduces failure probability by orders of magnitude in low‐error regimes, despite the increased number of sub‐computations introduced by the Ozaki decomposition. Furthermore, the overhead of ABFT remains small compared to the total number of photonic tensor‐core invocations, enabling high‐precision computation across a wide range of matrix sizes.
- 10 August 2026
- Published by Wiley in Concurrency and Computation: Practice and Experience
- Vol. 38 (16) , e70895
Open AccessReproducible Ozaki‐Based Matrix Multiplication on Photonic Matrix Processors With Algorithm‐Based Fault Tolerance
Photonic matrix processors are attracting attention in applications such as AI inference due to their low‐power and low‐latency properties. However, their applicability has been largely limited to low‐precision computation due to inherent stochastic analog noise and device non‐idealities. This limitation creates a large gap between the capabilities of photonic matrix processors and the potential demand for high‐precision computation in scientific computing. In this study, we propose a method to reproduce the results of the Ozaki‐based high‐precision matrix multiplication on low‐precision photonic matrix processors by combining the Ozaki scheme with a simple Algorithm‐Based Fault Tolerance (ABFT). The Ozaki scheme decomposes high‐precision operations into multiple low‐precision computations, while the ABFT introduces algorithm‐level redundancy to detect and correct errors caused by stochastic analog noise. Evaluation results show that ABFT significantly reduces failure probability by orders of magnitude in low‐error regimes, despite the increased number of sub‐computations introduced by the Ozaki decomposition. Furthermore, the overhead of ABFT remains small compared to the total number of photonic tensor‐core invocations, enabling high‐precision computation across a wide range of matrix sizes.
- Open Access
- 2 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. 37 (4) , 836-848
A 590-Nanosecond 757-Gbps FPGA Lossy Compressed Network
Inter-FPGA communication bandwidth has become a limiting factor in scaling memory-intensive workloads on FPGA-based systems. While modern FPGAs integrate high-bandwidth memory (HBM) to increase local memory throughput, network interfaces often lag behind, creating an imbalance between computation and communication resources. Data compression is a technique to increase effective communication bandwidth by reducing the amount of data transferred, but existing solutions struggle to meet the performance and operation latency requirements of FPGA-based platforms. This paper presents a high-throughput lossy compression framework that enables sub-microsecond latency communication in FPGA clusters. The proposed design addresses the challenge of aligning variable-length compressed data with fixed-width network channels by using transpose circuits, memory-bank reordering, and word-wise operations. A run-length encoding scheme with bounded error is employed to compress floating-point and fixed-point data without relying on complex fine-grained bit-level manipulations, enabling low-latency and scalable implementation. The proposed architecture is implemented on a custom Stratix 10 MX2100 FPGA card equipped with eight 50 Gbps network ports and silicon photonics transceivers. The system achieves up to 757 Gbps of aggregate bandwidth per FPGA in collective communication operations. Compression and decompression are performed within 590 ns total latency, while maintaining the quality of results in a GradAllReduce workload for deep learning.
- 2 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. 37 (4) , 836-848
Open AccessA 590-Nanosecond 757-Gbps FPGA Lossy Compressed Network
Inter-FPGA communication bandwidth has become a limiting factor in scaling memory-intensive workloads on FPGA-based systems. While modern FPGAs integrate high-bandwidth memory (HBM) to increase local memory throughput, network interfaces often lag behind, creating an imbalance between computation and communication resources. Data compression is a technique to increase effective communication bandwidth by reducing the amount of data transferred, but existing solutions struggle to meet the performance and operation latency requirements of FPGA-based platforms. This paper presents a high-throughput lossy compression framework that enables sub-microsecond latency communication in FPGA clusters. The proposed design addresses the challenge of aligning variable-length compressed data with fixed-width network channels by using transpose circuits, memory-bank reordering, and word-wise operations. A run-length encoding scheme with bounded error is employed to compress floating-point and fixed-point data without relying on complex fine-grained bit-level manipulations, enabling low-latency and scalable implementation. The proposed architecture is implemented on a custom Stratix 10 MX2100 FPGA card equipped with eight 50 Gbps network ports and silicon photonics transceivers. The system achieves up to 757 Gbps of aggregate bandwidth per FPGA in collective communication operations. Compression and decompression are performed within 590 ns total latency, while maintaining the quality of results in a GradAllReduce workload for deep learning.
- Open Access
- 11 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 211981-211995
Photonic Ising Machine Using Parallel Bandit Architecture Based on Laser Chaos
Photonic computing has emerged as a transformative technology for overcoming the fundamental limitations of semiconductor technology. Among recent advancements, the laser chaos decision maker has been introduced for ultrafast decision making by leveraging chaotic time series, primarily to solve the multi-armed bandit problem. In this study, we propose applying a laser chaos decision maker to a fully-connected Ising machine. Our method extends the applicability of the laser chaos decision maker, and achieves a first-ever Ising machine that obtains high-quality soloutions on the order of nanoseconds. Our approach addresses two critical challenges: the vast search space and the design of an algorithm to map the multi-armed bandit problem to Ising problems. To address these issues, we employ a parallel bandit architecture and design an algorithm to set the ground truth of decision-making repeatedly. Our GPU-based simulation evaluates the performance of our method using various benchmark problems and time series. We also analyze the effect of autocorrelation and the distribution of time series used by the laser chaos decision maker.
- 11 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 211981-211995
Open AccessPhotonic Ising Machine Using Parallel Bandit Architecture Based on Laser Chaos
Photonic computing has emerged as a transformative technology for overcoming the fundamental limitations of semiconductor technology. Among recent advancements, the laser chaos decision maker has been introduced for ultrafast decision making by leveraging chaotic time series, primarily to solve the multi-armed bandit problem. In this study, we propose applying a laser chaos decision maker to a fully-connected Ising machine. Our method extends the applicability of the laser chaos decision maker, and achieves a first-ever Ising machine that obtains high-quality soloutions on the order of nanoseconds. Our approach addresses two critical challenges: the vast search space and the design of an algorithm to map the multi-armed bandit problem to Ising problems. To address these issues, we employ a parallel bandit architecture and design an algorithm to set the ground truth of decision-making repeatedly. Our GPU-based simulation evaluates the performance of our method using various benchmark problems and time series. We also analyze the effect of autocorrelation and the distribution of time series used by the laser chaos decision maker.
- Open Access
- 19 February 2025
- Published by Association for Computing Machinery (ACM)
- p. 162-172
Lossy Compressed Collective Inter-FPGA Communications
- 19 February 2025
- Published by Association for Computing Machinery (ACM)
- p. 162-172
Open AccessLossy Compressed Collective Inter-FPGA Communications
- 7 November 2024
- Published by Optica Publishing Group in Journal of Optical Communications and Networking
- Vol. 17 (1) , A34-A45
In-network stable radix sorter using many FPGAs with high-bandwidth photonics [Invited]
A modern field-programmable gate array (FPGA) card can be equipped with high-bandwidth memory and high-bandwidth optical interconnection networks. This paper presents an in-network stable radix sorter on an eight-FPGA cluster. Each custom Stratix10 MX2100 FPGA card has up to 819-Gbps memory bandwidth (51.2Gbps×16channels) and up to 800-Gbps network bandwidth (25Gbps×32channels) with eight custom embedded optical modules. Our radix sorter for a 32-bit key range consists of eight 4-bit counting sorts optimized to in-network processing. Each counting sort needs only one memory read/write access for improving its throughput. We demonstrated a sorting throughput of 37.2 GB/s and an energy efficiency of 9.2 MB/J for 32-bit key range and 16-GiB data size using eight memory channels with 409.6 Gbps memory bandwidth per FPGA. It can scale up to 256 FPGAs with a sorting throughput of 983 GB/s for a 32-bit key range and 512-GiB data size.
- 7 November 2024
- Published by Optica Publishing Group in Journal of Optical Communications and Networking
- Vol. 17 (1) , A34-A45
In-network stable radix sorter using many FPGAs with high-bandwidth photonics [Invited]
A modern field-programmable gate array (FPGA) card can be equipped with high-bandwidth memory and high-bandwidth optical interconnection networks. This paper presents an in-network stable radix sorter on an eight-FPGA cluster. Each custom Stratix10 MX2100 FPGA card has up to 819-Gbps memory bandwidth (51.2Gbps×16channels) and up to 800-Gbps network bandwidth (25Gbps×32channels) with eight custom embedded optical modules. Our radix sorter for a 32-bit key range consists of eight 4-bit counting sorts optimized to in-network processing. Each counting sort needs only one memory read/write access for improving its throughput. We demonstrated a sorting throughput of 37.2 GB/s and an energy efficiency of 9.2 MB/J for 32-bit key range and 16-GiB data size using eight memory channels with 409.6 Gbps memory bandwidth per FPGA. It can scale up to 256 FPGAs with a sorting throughput of 983 GB/s for a 32-bit key range and 512-GiB data size.
- Open Access
Performance Improvement of Federated Learning Server using Smart NIC
Federated learning is a distributed machine learning approach where local weight parameters trained by clients locally are aggregated as global parameters by a server. The global parameters can be trained without uploading privacy-sensitive raw data owned by clients to the server. The aggregation on the server is simply done by averaging the local weight parameters, so it is an I/O intensive task where a network processing accounts for a large portion compared to the computation. The network processing workload further increases as the number of clients increases. To mitigate the network processing workload, in this paper, the federated learning server is offloaded to NVIDIA BlueField-2 DPU which is a smart NIC (Network Interface Card) that has eight processing cores. Dedicated processing cores are assigned by DPDK (Data Plane Development Kit) for receiving the local weight parameters and sending the global parameters. The aggregation task is parallelized by exploiting multiple cores available on the DPU. To further improve the performance, an approximated design that eliminates an exclusive access control between the computation threads is also implemented. Evaluation results show that the proposed DPDK-based federated learning server on the DPU with the approximation accelerates the execution time by 1.39 times with a negligible accuracy loss compared with a baseline server on the host CPU.
Open AccessPerformance Improvement of Federated Learning Server using Smart NIC
Federated learning is a distributed machine learning approach where local weight parameters trained by clients locally are aggregated as global parameters by a server. The global parameters can be trained without uploading privacy-sensitive raw data owned by clients to the server. The aggregation on the server is simply done by averaging the local weight parameters, so it is an I/O intensive task where a network processing accounts for a large portion compared to the computation. The network processing workload further increases as the number of clients increases. To mitigate the network processing workload, in this paper, the federated learning server is offloaded to NVIDIA BlueField-2 DPU which is a smart NIC (Network Interface Card) that has eight processing cores. Dedicated processing cores are assigned by DPDK (Data Plane Development Kit) for receiving the local weight parameters and sending the global parameters. The aggregation task is parallelized by exploiting multiple cores available on the DPU. To further improve the performance, an approximated design that eliminates an exclusive access control between the computation threads is also implemented. Evaluation results show that the proposed DPDK-based federated learning server on the DPU with the approximation accelerates the execution time by 1.39 times with a negligible accuracy loss compared with a baseline server on the host CPU.
- Open Access
- 28 June 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 65590-65598
Experimental Demonstration of Approximate Communication Based on Radio-Over- Fiber Systems
The importance of multi-valued data transmission for improving communication throughput has become apparent with the explosive growth o communication and computing demands. Although modulation schemes with denser constellations can provide increased data rates, they require a higher signal-to-noise ratio. On the other hand, sparser constellation formats offer reliable communication thanks to their greater tolerance to noise, but the throughput is inferior to that with denser constellations. Here, we introduce an approximate computing/communication that involves brings the critical notion of quality of information to tolerate the trade-off between throughput and reliability, which will be important in future high-performance optical communications and interconnects. In this study, we experimentally demonstrate a dynamically configured transmission scheme where different modulation formats are adopted according to the importance of bit sequences on the basis of Radio-over-Fiber (RoF) systems in transmitting floating-point numbers. The proposed system demonstrated increased gross and net throughput, while errors in the transmitted floating-point numbers were uniformly bounded.
- 28 June 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 65590-65598
Open AccessExperimental Demonstration of Approximate Communication Based on Radio-Over- Fiber Systems
The importance of multi-valued data transmission for improving communication throughput has become apparent with the explosive growth o communication and computing demands. Although modulation schemes with denser constellations can provide increased data rates, they require a higher signal-to-noise ratio. On the other hand, sparser constellation formats offer reliable communication thanks to their greater tolerance to noise, but the throughput is inferior to that with denser constellations. Here, we introduce an approximate computing/communication that involves brings the critical notion of quality of information to tolerate the trade-off between throughput and reliability, which will be important in future high-performance optical communications and interconnects. In this study, we experimentally demonstrate a dynamically configured transmission scheme where different modulation formats are adopted according to the importance of bit sequences on the basis of Radio-over-Fiber (RoF) systems in transmitting floating-point numbers. The proposed system demonstrated increased gross and net throughput, while errors in the transmitted floating-point numbers were uniformly bounded.
- 8 April 2023
- Published by Springer Nature
- p. 401-412
A High-Radix Circulant Network Topology for Efficient Collective Communication
- 8 April 2023
- Published by Springer Nature
- p. 401-412
A High-Radix Circulant Network Topology for Efficient Collective Communication
- 8 April 2023
- Published by Springer Nature
- p. 301-313
A Hardware Trojan Exploiting Coherence Protocol on NoCs
- 8 April 2023
- Published by Springer Nature
- p. 301-313
A Hardware Trojan Exploiting Coherence Protocol on NoCs