All times are in CT (local time in Chicago)
Tuesday, 17 November 2026
1:30 PM - 2:00 PM
Mixed-Precision Cholesky Factorization for Semidefinite Programming Using Ozaki Scheme II
Haruki Nakamae
School of Computing, Institute of Science Tokyo
This talk explores the accuracy and performance of a mixed-precision Cholesky factorization based on the INT8 Ozaki scheme II. Within the tiled Cholesky algorithm, we apply an optimized Ozaki scheme II to the symmetric rank-k update (SYRK), while keeping the diagonal factorization (POTRF) and the triangular solve (TRSM) in double precision. We further introduce a look-ahead technique and a dynamic tile size scheduler. On a server whose GPU has limited FP64 throughput, our implementation attains a 3.04x speedup over the OpenBLAS dpotrf routine while retaining double-precision accuracy. We also examine how the number of slices affects the backward error and the relative forward error, and find that 15 slices are required to preserve double-precision accuracy within the range covered by our experiment. Finally, we integrate the proposed factorization into a double-precision semidefinite programming solver and obtain a 1.47x speedup on an SDP benchmark instance (1et.2048).
2:15 PM - 2:45 PM
Evaluation of Large-Scale Supercomputer Network Topology Using Betweenness Centrality
Toshio Endo
Supercomputing Research Center, IIR, Institute of Science Tokyo
In large-scale supercomputers, efficient communication among compute nodes is crucial for the performance of parallel computing. To improve communication performance, many network topologies have been proposed and continue to be explored. To accelerate topology exploration, we present an efficient evaluation method of network topologies based on graph theory. Rather than latency metrics, we focus on betweenness centrality, a bandwidth-centric metric. We discuss trade-offs among full fat-tree, tapered fat-tree, and dragonfly.
3:00 PM - 3:30 PM
Proxy-surface-based fast direct solver for TE-mode scattering problems on distributed memory systems
Yasuhiro Matsumoto
Center for Information Infrastructure, Institute of Science Tokyo
This talk presents an MPI/OpenMP hybrid fast direct solver for TE-mode electromagnetic scattering, formulated as a two-dimensional Helmholtz problem. To overcome the limited accuracy of conventional HSS solvers due to a support of local corrected quadratures, we employ the weakly singular Burton–Miller boundary integral equation and a Nyström method with one-point correction, achieving O(h^3) convergence instead of the typical O(h). We also introduce a load-balancing method to improve parallel efficiency, since matrix component calculation dominates computational costs. Numerical results demonstrate high accuracy and nearly ideal strong and weak scalability on TSUBAME4.0.
Wednesday, 18 November 2026
2:00 PM - 2:30 PM
Nation-wide GPU Application Development in Japan by Agentic-AI
Taisuke Boku
Director of Advanced HPC-AI Research and Development Support Center (HAIRDESC)
Code change from CPU-oriented scientific applications to GPU-ready ones is the most urgent issue in supercomputing in Japan, especially for HPCI (HPC Infrastructure) systems with advanced GPU, including FugakuNEXT. Japanese government launched a special program to solve this issue in 2025, and HAIRDESC (Advanced HPC-AI Research and Development Support Center) was established for GPU programming education and human resource development, code shift from CPU to GPU on nation-wide applications, and promotion of GPU computing for all scientists. The program will continue to FY 2029, until FugakuNEXT system is deployed as well as most of advanved supercomputers under HPCI organizations.
Converting existing application codes developed for CPU-based supercomputers to run on GPUs requires an enormous amount of human effort. It is virtually impossible to complete the porting of several dozen government-supported applications within a few years through manual efforts alone. To address this challenge, we are working to achieve this goal by fully embracing AI, particularly Agentic-AI, in close collaboration with the original application developers. Current AI systems have become capable of generating code of sufficiently high quality even through relatively inexpensive subscription plans. At HAIRDESC, we are developing a harness to apply these AI capabilities to a wide range of HPC applications.
Ultimately, we plan to establish a set of skills and procedures that can be applied to various HPC applications in Japan, primarily targeting applications developed under projects selected for the Grand Reach Program of Japan's Ministry of Education, Culture, Sports, Science and Technology (MEXT).
In this talk, I will present our activities toward these goals.