In the era of modern science, scientists have developed numerical models to predict and understand the weather and ocean phenomena based on fluid dynamics. While these models have shown high accuracy at kilometer scales, they are operated with massive computer resources because of their computational complexity. In recent years, new approaches to solve these models based on machine learning have been put forward. The results suggested that it be possible to reduce the computational complexity by Neural Networks (NNs) instead of classical numerical simulations. In this project, we aim to shed light upon different ways to accelerating physical models using NNs. We test two approaches: Data-Driven Statistical Model (DDSM) and Hybrid Physical-Statistical Model (HPSM) and compare their performance to the classical Process-Driven Physical Model (PDPM). DDSM emulates the physical model by a NN. The HPSM, also known as super-resolution, uses a low-resolution version of the physical model and maps its outputs to the original high-resolution domain via a NN. To evaluate these two methods, we performed idealized experiments with a quasi-geostrophic model and measured their accuracy and their computation time. The results show that HPSM reduces the computation time by a factor of 3 and it is capable to predict the output of the physical model at high accuracy up to 9.25 days. The DDSM reduces the computation time even further by a factor of 4, but its predictability is limited to at most only within 2 days. These first results are promising and imply the possibility of bringing complex physical models into real time systems with lower-cost computer resources in the future.

Important Notes:
- Please turn off your video and microphone when you join the meeting. - The broadcasting may be interrupted or terminated depending on the network condition or any other unexpected event. - The program schedule and contents may be modified without prior notice. - Depending on the utilized device and network environment, it may not be able to watch the session. - All rights concerning the broadcasted material will belong to the organizer and the presenters, and it is prohibited to copy, modify, or redistribute the total or a part of the broadcasted material without the previous permission of RIKEN.

In order to analyze the functions and interactions of biomolecules at the molecular level, it is necessary to elucidate the detailed three-dimensional (3D) structure of biomolecules. Single-particle 3D structure analysis using X-ray free electron laser (XFEL) is a new structural biology technique that allows observation of biomolecular complexes and tissues that are difficult to crystallize in a state close to nature. However, even using intensive XFEL beam, diffraction intensity obtained from a single biomolecule is still weak, and the S/N ratio of the diffraction pattern is small. In order to restore the 3D structure with the expected resolution, it is necessary to estimate the requirement for the quantity and quality of the diffraction patterns. For this purpose, we tried to reconstruct the 3D structure of ribosome using the simulation diffraction pattern sets, consisting of the different number of patterns created by different beam intensities. In this presentation, I will show how the resolution of the reconstructed 3D structure depends on the quantity and quality of the diffraction patterns, which indicate the minimum requirement of XFEL experimental conditions to obtain the diffraction patterns to reconstruct the molecular structure with several nanometers resolution. Also, I will introduce our recent results of restoration of 3D-structure of gold nano-particle from the diffraction patterns obtained by XFEL experiment.

Important Notes:
- Please turn off your video and microphone when you join the meeting. - The broadcasting may be interrupted or terminated depending on the network condition or any other unexpected event. - The program schedule and contents may be modified without prior notice. - Depending on the utilized device and network environment, it may not be able to watch the session. - All rights concerning the broadcasted material will belong to the organizer and the presenters, and it is prohibited to copy, modify, or redistribute the total or a part of the broadcasted material without the previous permission of RIKEN.

As a trend in computer architecture, "throughput oriented architecture" is widely adopted to increase throughput performance by duplicating computing resources. Recent improvements in the performance of general purpose processors are also due to the expansion of the data width of instruction sets, the introduction of multiple cores and fast memory hierarchies, which moves closer to typical throughput oriented architectures such as Graphics Processing Unit (GPU). In this talk, we apply the OpenCL throughput oriented programming model to the Fugaku supercomputer to exploit potential performance of hierarchical computing resources. The compiler translates OpenCL kernel code into ARM SVE SIMD code with thread and NUMA runtime. The overview of the code translation from OpenCL to parallel code will be given in the presentation.

Important Notes:
- Please turn off your video and microphone when you join the meeting. - The broadcasting may be interrupted or terminated depending on the network condition or any other unexpected event. - The program schedule and contents may be modified without prior notice. - Depending on the utilized device and network environment, it may not be able to watch the session. - All rights concerning the broadcasted material will belong to the organizer and the presenters, and it is prohibited to copy, modify, or redistribute the total or a part of the broadcasted material without the previous permission of RIKEN.

In modern lattice QCD (LQCD) computations, where physical light quarks in the standard model of particles are simulated directly, the fermion matrix - linear solver consumes most of the CPU time along the simulations. The QCD wide SIMD (QWS) library, which has been made in the LQCD co-design activity implementing the linear solver, makes it possible to efficiently use the power of Fugaku in the state-of-the-art LQCD simulations. On the other hand there is an international activity on the code development with a different approach to utilize SIMD. The Japanese LQCD community is making use of these two products depending on the problem to solve. The current status of these activities as well as an outlook on the impact to the particle/nuclear physics are discussed.

Important Notes:
- Please turn off your video and microphone when you join the meeting. - The broadcasting may be interrupted or terminated depending on the network condition or any other unexpected event. - The program schedule and contents may be modified without prior notice. - Depending on the utilized device and network environment, it may not be able to watch the session. - All rights concerning the broadcasted material will belong to the organizer and the presenters, and it is prohibited to copy, modify, or redistribute the total or a part of the broadcasted material without the previous permission of RIKEN.

For the HPC community, common issues on computational speed and computational accuracy are generally considered to be conflicting. However, the diversity and enhancement of hardware and the high productivity of software have allowed users to choose the precision within the requirement of appropriate computational accuracy. These may provide us with enormous changes in scientific and technical computing, whereas it has been dominated by double-precision calculation for a long time. In the seminar, I will introduce the recent topics such as the relationship between high performance and precision and the relationship between modern hardware and computation accuracy, mainly focusing on the numerical libraries developed by my team in the above topics; i) establishment of higher precision software by massively-and-high-performance low-precision computing units, ii) algorithmic advancement of lower-precision units in scientific computing like HPL-AI benchmark, iii) idea of minimal-precision computing. The first is the realization of a DGEMM-equivalent matrix product using TensorCore(TC) by Mukunoki et al. This is an important fact. It is one of the academic case studies of the utilization of TC's. On the other hand, it suggests the possibility of controlling the number of double precision units by installing a sufficient number of low precision arithmetic units. The second refers to our HPL-AI result, of course, one of the world's four crowning benchmarks and its computation is based on a mixed precision of FP16, FP32, and FP64 formats. The essential point of HPL-AI is to bring out the high performance of low-precision arithmetic while preventing numerical instability and inaccuracy in low-precision arithmetic. It is not simply a matter of rewriting double to half. This is accomplished by a preliminary analysis of the computation target and patterns. The third is to promote the minimum system of computation, which is anticipated to change storage capacity, energy consumption, and minimum hardware requirements of the current floating-point unit. Users won't feel a big impact in terms of input/output, but the internal design of computers will be significantly enhanced.

Important Notes:
- Please turn off your video and microphone when you join the meeting. - The broadcasting may be interrupted or terminated depending on the network condition or any other unexpected event. - The program schedule and contents may be modified without prior notice. - Depending on the utilized device and network environment, it may not be able to watch the session. - All rights concerning the broadcasted material will belong to the organizer and the presenters, and it is prohibited to copy, modify, or redistribute the total or a part of the broadcasted material without the previous permission of RIKEN.

The High Performance Big Data Research Team at RIKEN Center for Computational Science (RIKEN R-CCS) has been researching and developing system software to facilitate extreme-scale big data processing, machine learning and deep learning for high performance computing (HPC) systems, i.e., convergence of AI/Big Data and HPC. In this talk, Kento Sato introduce several R&D activities for accelerating AI/Big data applications on HPC systems (HPC for AI/Big data) as well as ones for resolving HPC challenges by using these AI/Big data techniques (AI/Big data for HPC).

Important Notes:
- Please turn off your video and microphone when you join the meeting. - The broadcasting may be interrupted or terminated depending on the network condition or any other unexpected event. - The program schedule and contents may be modified without prior notice. - Depending on the utilized device and network environment, it may not be able to watch the session. - All rights concerning the broadcasted material will belong to the organizer and the presenters, and it is prohibited to copy, modify, or redistribute the total or a part of the broadcasted material without the previous permission of RIKEN.

A whole-brain simulation allows us to investigate all interactions among neurons in the brain to understand the mechanisms of information processing and brain diseases. However, the whole-brain simulation has not been realized so far due to a lack of physiological data and computational resources. The supercomputer Fugaku is estimated to perform the whole-brain simulation on a human-scale due to its computational performance. To realize the whole-brain simulation and understand the brain's mechanism, we have worked in Post-K exploratory challenge 4-1 and started a new project of the Fugaku since April 2020. In the projects, we constructed a minimum unit of mammalian brain model consisting of the cerebral cortex, basal ganglia, cerebellum, and thalamus, which reproduced resting-state neural activity. We also have been developing an in-house neural network simulator, MONET, that can utilize the wide SIMD and many cores of A64FX of the Fugaku. The MONET simulator simulated models of the cerebral cortex with 5 billion neurons and cerebellum with 64 billion on the K computer. A recent experiment showed that MONET realized 40 times the more excellent computational performance per compute nodes for simulations of the cerebral cortex and cerebellum on the Fugaku evaluation environment III than the K computer. These results suggest that the full system of the Fugaku may realize whole-brain simulation and contribute to an understanding of the brain.

Important Notes:
- Please turn off your video and microphone when you join the meeting. - The broadcasting may be interrupted or terminated depending on the network condition or any other unexpected event. - The program schedule and contents may be modified without prior notice. - Depending on the utilized device and network environment, it may not be able to watch the session. - All rights concerning the broadcasted material will belong to the organizer and the presenters, and it is prohibited to copy, modify, or redistribute the total or a part of the broadcasted material without the previous permission of RIKEN.

We use floating-point numbers for numerical computations in computers. However, as floating-point numbers are represented in finite-precision (usually, 32-bit or 64-bit), the computation has rounding errors. The accumulation may result in inaccuracy of the computation. Moreover, the result can be non-identical at the rounding-error level on different computational environments because the rounding-errors are introduced depending on the order of the computation. Our team studies and develops technologies for addressing those issues to improve the reliability of numerical computations. In this talk, I will introduce several studies that I have been conducting now, including (1) high-precision matrix-multiplications using low-precision operations, (2) accurate and reproducible BLAS and sparse linear solvers, and (3) minimal-precision computing project.

Important Notes:
- Please turn off your video and microphone when you join the meeting. - The broadcasting may be interrupted or terminated depending on the network condition or any other unexpected event. - The program schedule and contents may be modified without prior notice. - Depending on the utilized device and network environment, it may not be able to watch the session. - All rights concerning the broadcasted material will belong to the organizer and the presenters, and it is prohibited to copy, modify, or redistribute the total or a part of the broadcasted material without the previous permission of RIKEN.

Atomic Force Microscopy (AFM) is an experimental technique to study structure-function relationship of biomolecules. AFM provides two-dimensional (2D) images of biomolecules at nanometer resolution, however information of three-dimensional (3D) structures is beneficial to understand conformational change often encoded in the images. Our aim is to recover 3D information from an AFM image by computational modeling. In our method we use Monte Carlo sampling, in which candidate models are generated in such a way that the similarity between pseudo-AFM images generated from the models and target AFM image increases gradually. First, we tested our algorithm theoretically on two proteins to model their conformational transitions. Using an AFM image as reference, the algorithm can produce a low-resolution 3D model of the target molecule. Now we are applying our method to experimentally obtained AFM data for a protein Clpb. This is a heat-shock protein with hexameric arrangement and its high-speed AFM images include at least four major classes of structures. So far, we recovered the candidate models corresponding to those structural classes enabling us to realize 3D representations encoded in these AFM images. The new AFM based hybrid modeling method would be useful to retrieve 3D information of a biomolecule, therefore we could study how 3D structures of biomolecules are dictating its function. Moreover, from the analysis of the reconstructed models it is possible to understand in more detail the conformational transitions of proteins observed in near physiological condition.

Important Notes:
- Please turn off your video and microphone when you join the meeting. - The broadcasting may be interrupted or terminated depending on the network condition or any other unexpected event. - The program schedule and contents may be modified without prior notice. - Depending on the utilized device and network environment, it may not be able to watch the session. - All rights concerning the broadcasted material will belong to the organizer and the presenters, and it is prohibited to copy, modify, or redistribute the total or a part of the broadcasted material without the previous permission of RIKEN.

The goal of the galaxy formation simulation is to follow the formation of galaxies through the whole history of the universe. To achieve this goal, the gravitational dynamics of dark matter, gas and stars, hydrodynamics of gas, star formation process, energy input from stars and supernovae to gas should be modeled properly. Traditionally, the individual timestep algorithm has been used to handle regions with short timescales, such as star-forming regions and supernova shells. However, this algorithm does not scale well for more than 10M particles, because of the need for global communication at each timestep. We have developed a new algorithm to handle regions with short timescales. We report the performance of our new code based on this new algorithm on Supercomputer Fugaku. It scales well for more than 10G particles and more than 100K cores. We also discuss our new DSL for the generation of efficient kernel for particle-particle interaction.

Important Notes:
- Please turn off your video and microphone when you join the meeting. - The broadcasting may be interrupted or terminated depending on the network condition or any other unexpected event. - The program schedule and contents may be modified without prior notice. - Depending on the utilized device and network environment, it may not be able to watch the session. - All rights concerning the broadcasted material will belong to the organizer and the presenters, and it is prohibited to copy, modify, or redistribute the total or a part of the broadcasted material without the previous permission of RIKEN.