Digital twins represent the next frontier in the impact of computational science on grand challenges across science, technology and society. A digital twin is a computational model or set of coupled models that evolves over time to persistently represent the structure, behavior, and context of a unique physical system, process or biological entity. Digital twins have the potential to enable safer and more efficient engineering systems, a greater understanding of the natural world around us, and better medical outcomes for all of us as individuals. This talk will highlight progress and opportunities in achieving robust, reliable digital twins at scale, including the important role of reduced-order modeling, scientific machine learning and uncertainty quantification.

We are on the verge of a global communications revolution based on ubiquitous high-speed 5G, 6G, and free-space optics technologies. The resulting global communications fabric can enable new ultra-collaborative research modalities that pool sensors, data, and computation with unprecedented flexibility and focus. But realizing these modalities requires new services to overcome the tremendous friction currently associated with any actions that traverse institutional boundaries. The solution, I argue, is new global science services to mediate between user intent and infrastructure realities. I describe our experiences building and operating such services and the principles that we have identified as needed for successful deployment and operations.

High-performance systems are increasingly bottlenecked by the energy and communications costs of interconnecting numerous compute and memory resources. Integrated silicon photonics offer the opportunity of embedding optical connectivity that directly delivers high off-chip communication bandwidth densities with low power consumption. Our recent work has shown how integrated silicon photonics with comb-driven dense wavelength-division multiplexing can scale to realize Pb/s chip escape bandwidths with sub-picojoule/bit energy consumption. Beyond alleviating the bandwidth/energy bottlenecks, embedded photonics can enable new architectures that leverage the distance independence of optical transmission with flexible connectivity tailored to accelerate distributed ML applications.

We’re going through a revolutionary change in computing right now. Software and hardware are radically adapting to new breakthroughs in artificial intelligence, which is itself driven by massive disruptions in the way we design hardware for processing data. Where are we going next and how do we deliver the foundations to achieve it? Andrew has been working in the field of new languages and compilers for new processors for the last 25 years. This talk will be his vision for the future: new languages, algorithms, applications and hardware and how we enable them with solid foundations.

So we have arrived in the Exascale era, now what? Even with the massive improvements in computing, many applications are dominated by data movement and cannot achieve close to the peak performance of the machine. As we innovate on the future architectures for HPC and AI, there are fundamental needs to improve the data movement aspect, and build a more balanced machine. In this talk, we will touch on some of the technologies that Samsung is innovating aiming at addressing the data movement bottleneck. We will also discuss the new paradigm of memory coupled compute, and how this new technology is different and will provide further advancement in breaking the memory and communication wall.

In May 2022 the HPE Cray “Frontier” system at Oak Ridge National Lab broke the Exascale barrier for the first time. This remarkable accomplishment culminates a decadal effort, and opens a new path for scientific discovery at scale for the next several years to come. Progress doesn’t stop at exascale, and we are entering a new era where different considerations will drive the next generation of technology directions. At HPE we have come up with “seven pillars” for post-exascale supercomputing: programmer productivity, edge-to-exascale orchestration, a self-learning and trustworthy AI data foundation, heterogeneity across the board, a data-accelerated storage infrastructure, a common network fabric infrastructure, and pervasive security. This talk will cover the technical vision, the individual technologies, the challenges, and will highlight opportunities for further research and development.

“Be careful what you wish for. You might get it.” I have been developing software for almost 50 years and supporting myself doing so for more than 45 of those years. Although I do occasionally code up something just for fun, the vast majority of the software that I have written has been requested (and paid for) by others. In other words, I have spent most of my career writing software that other people asked for. As many have observed, it is hard enough to write software that runs correctly, but even harder to work out the correct software to write. This talk will tell the tale of a few of my attempts to correctly write the correct software.

Following up on the success of Fugaku, Riken R-CCS is now on course to realize its successor, Fugaku-NEXT, ideally by the end of the 2020s. However, the next generation breakthrough in performance is being inhibited by a variety of factors pertaining to the slowdown of Moore’s Law. In fact, such difficulty is generating a series of rather technically unfounded views on evolutionary paths on computing, or ‘myths’, that are distorting the right way of moving forward. As a institute of leading edge science, we are undertaking a scientific, methodological approach to how next generation machines may achieve more than an order of magnitude performance gains over Fugaku, while retaining its other trait, being broad and general purpose to be applicable to wider ranging problems we face today as a society today.

With the recent arrival of the Frontier system in the US at Oak Ridge National Laboratory (ORNL), and with the applications and software technologies under development as part of the US Department of Energy (DOE) Exascale Computing Project (ECP) now poised to exploit Frontier’s capabilities to tackle problems of national and international interest, the highly anticipated “dawn of the exascale computing era” is here. This is indeed a very exciting time for the world’s high-performance computing (HPC) community, as many exascale uncertainties and challenges over the past decade or so have been surpassed. Given the concerted US DOE investments in the ECP and architectural co-design embodied in Frontier, this exascale system is and will continue to be “used, useful, and affordable” over its lifetime. Exascale-capable applications are a foundational element of the ECP and the vehicle for delivery of mission need on targeted exascale systems such as Frontier. The ECP’s mission need application projects, each addressing an exascale challenge problem—a high-priority strategic problem of national interest that is intractable without at least 50 times the computational power of the HPC systems available at the project’s inception in 2016. Exascale applications are built on underlying software technologies, which play an essential supporting role in application efficacy on computing systems. The ECP’s ST effort is developing an expanded and vertically integrated software stack that includes advanced mathematical libraries, extreme-scale programming environments, development tools, visualization libraries, and the software infrastructure to support large-scale data management and data science for science and security applications. The ST efforts complement and integrate into the broader scientific software ecosystem that includes capabilities from industry and the broader HPC R&D community. Architectural details of the Frontier system will be given along with the challenges overcome in readying traditional and new exascale software technologies and applications as part of the ECP.

Urgent science describes time-critical, data-driven scientific workflows that can leverage distributed data sources in a timely way to facilitate important decision making. Despite the exponential growth of available digital data sources and the ubiquity of non-trivial computational power for processing this data, realizing such urgent science workflows remains challenging. In this talk I will explore how the computing continuum, spanning resources at the edges, in the core and in-between, can be harnessed to support urgent science. I will also describe recent research in programming abstractions that can express what data should be processed and when and where it should be processed, middleware services that automate the discovery of resources and the orchestration of computations across these resources.

OpenMP has established itself as the standard for shared memory parallel programming in HPC. One of the main advantages is the ease of use. A simple pragma added to the source code is sufficient to parallelize a block of code. All the complexity behind this, is hidden in the compiler and run time library. This ease of use also has a somewhat darker side though. As easy as it is to parallelize code, it is also just as easy to make a mistake that affects the correctness of results, or the performance. In this talk, we start with a very quick overview of OpenMP and explain the concepts needed in the remainder of this talk. This is followed by some common pitfalls and the solution. We end with a case study that shows how a targeted approach and tuning strategy led to a dramatic improvement in the performance of an application from graph analysis.

Scientific workflows are now a common tool used by domain scientists in a number of disciplines. They are appealing because they enable users to think at a high level of abstraction, composing complex applications from individual application components. This talk focuses on the Pegasus workflow management system, which automates the process of executing workflows on modern cyberinfrastructure. It takes high-level, resource-independent descriptions and maps them onto the available heterogeneous resources: campus clusters, high-performance computing resources, high-throughput resources, clouds, and now the edge. This talk will describe the key concepts used in the Pegasus WMS and how they allowed the system to incorporate the edge as a computing platform for scientific workflows.

The Square Kilometer Array Telescope was just begun construction of phase one, with 5 years to deliver two telescopes, each with a collecting area of just 10% of the full million square meters of antenna collecting area. Even at this limited scale the computing challenges are significant. I will talk about the size of the problem, how the design is partitioned, so that it can be scaled through the delivery milestones. We will explore some problems in the transport of the huge volumes of data, and some techniques I am proposing to alleviate the pressure.

Future high-performance computing (HPC) challenges will be driven by the convergence of physical simulation, artificial intelligence & machine learning (AI/ML), and data science computing capabilities. While computational performance gains afforded by Moore’s Law have enabled large-scale HPC system design and deployment using commodity CPU and GPU processing components, new near and long-term technologies will be required to effectively support such converged workloads. These new technologies will integrate commodity computing components with custom domain-specific accelerators into ever-more heterogeneous architectures resulting in a diverse ecosystem of industry technology developers, university, and U.S. Government researchers. This overview describes efforts to construct an end-to-end co-design framework that lays a groundwork for such an ecosystem, including notable outcomes, remaining challenges, and future opportunities.

Transformers have become the dominant neural network architecture not only in natural language processing, but also in computer vision and other modalities. The potential of transformers lies in its scaling laws, where pre-training larger models on larger datasets leads not only to increased accuracy on a wide range of downstream tasks, but also emergence of new capabilities. However, the need for very large datasets poses many challenges, which include societal biases, copyright, and privacy in data scraped from the Internet. The cost to clean the datasets to avoid these issues becomes prohibitive at scale, which is the next major challenge in deep learning. In this talk, I will discuss the possibility of using synthetic datasets that are free of such issues.

In the first part of this talk talk we discuss the concept of extreme performance HPC where the applications often include latencies below 100 ms or even below 10 ms. To facilitate low latency computation has to be placed in the vicinity of the end users by utilizing the concept of Edge Computing. We present the novel failure resilience mechanisms applied to Edge systems considering timeliness, hyper-heterogeneity and resource scarcity. We discuss our machine learning based mechanism that evaluates the failure resilience of a service deployed redundantly on the Edge infrastructures. Our approach learns the spatiotemporal dependencies between Edge server failures and combines them with the topological information to incorporate link failures by utilizing the concept of the Dynamic Bayesian Networks (DBNs). Eventually, we infer the probability that a certain set of servers fails or disconnects concurrently during service runtime. The robustness of the approach has been evaluated using the concept of code offloading. In the second part we discuss the potential impact of Edge Computing in the current rise of highly specialized architectures ranging from neuromorphic to quantum computing. As we experience the paradigm shift from generalized architectures of the Von Neumann era to highly specialized architectures in the Post-Moore’s law era we expect the coexistence of multiple types of architectures specialized for different types of computation. We define the two of the architectures that attracted the most interest in the research community, where we witness not only theoretical developments but also first implementations and practical use cases. Afterward, we discuss the first ideas but also challenges in the integration of identified architectures into existing HPC systems.

How to structure the object and detection problem with high resolution images to leverage the scale of economies of devices, high resolution cameras, 5G radio, CPU, GPU and neural compute. It would be awesome to share some of the work we have been doing. The two main areas which people might be interested in are; High performance object detection using GPU/TPU on smartphones. Computing 3D surface geometry for shiny subjects like polished metal parts for inspection – along with a CUDA stack for processing normal maps rather than greyscale or RGB images. As a third item I also have the Quest Pro and I could show some of the real time VR/AR simulations we are rendering.

A Digital Twin is so much much more than a visual experience or a digital simulation of a specific real-world process. It can be an immersive data visualization and understanding experience, a critical foundation for the application and communication of AI and simulation tools. It can be delivered on mobile, desktop, AR/VR to deliver value for decades or centuries. It can help make an organization cleaner, greener, safer and more resource optimized. It can understand and represent complex supply chains and geospatial and scanned data. However these capabilities depend on the ability to assimilate data from many sources and standards and amalgamate it into standardized structures suitable for simulating and visualizing in whatever specialized systems are required. It is the rigorous application of the concept of Ontology principles that allow this to happen. This will be discussed and presented with examples.

DOE has just deployed its first Exascale system at ORNL, so now is an appropriate time to revisit our Exascale predictions from over a decade ago and think about post-Exascale. We are now seeing a Cambrian explosion of new technologies during this ‘golden age of architectures,’ making codesign of architectures with software and applications more critical than ever. In this talk, I will revisit Exascale trajectory, survey post-Exascale technologies, and discuss their implications for both system design and software. As an example, I will describe Abisko, a new microelectronics codesign project, that focuses on designing a chiplet for analog spiking neural networks using novel neuromorphic materials.

Recent advances in co-packaged optics make it possible to drive multiple terabytes per second out of a single socket. The photonic eco-system is advancing rapidly, making co-packaged optics an excellent candidate for upcoming generations of post-exascale systems, bringing speed-of-light latency for short- and long-reach communication, combined with an exponential growth of communication bandwidth and connectivity: it is now possible to drive 32-64 optical connections out of a single high-radix device. This level of connectivity is a real advancement, but current network designs do not fully exploit this opportunity. Without advances in system design, these new systems will not reach their potential. Such advances are especially needed in network topology and system design, which are still open areas of research. This talk is on the border of mathematics and computer science, working toward the design of such a topology. We will discuss the use of mathematical graph theory and projective geometry in the design of very large and compact interconnection networks that are optimally tailored to this emerging technology. We have used classical graph theory and known graphs to create PolarFly, a new family of diameter-2 topologies. This topology supports radixes suited to the new high-radix routers, asymptotically approaches the maximum number of nodes for the radix and diameter, exploits mathematical symmetries for modularity, and outperforms other networks in terms of scalability, cost and performance. Diameter 2 is suited to smaller systems, but not exascale. We will discuss current and future directions as well, aimed at exascale and post-exascale systems.

AI, as seen today by large-scale deep learning, presents complex challenges to computer systems. It requires the adaptive execution of heterogeneous components, each of which is a cluster of parallel tasks. Further, large amounts of data need to be read, written, and often moved in small dynamic batches. Deep learning must be efficiently linked to pre and post-processing (data engineering). This implies converged environments with Java, Python, and C++ ecosystems. Further, AI must be integrated with better-understood large parallel simulations. The simulations can invoke AI to control passage through phase space (often with ensembles). Alternatively, it can train surrogates used to replace all or part of the simulation. In the latter case, there are consequent inferences, as in computational fluid dynamics or climate simulations where AI can learn microstructure. This implies we must support systems that run well in conjunction with classic Slurm-MPI-based simulations as well as in modern cloud environments, including challenges from shared resources due to multi-tenancy. This extends the needed convergence to link HPC and cloud environments. The talk details these challenges and describes a possible approach

Influenced by the advances in data and computing, the scientific practice increasingly involves machine learning and artificial intelligence (AI) driven methods which require specialized capabilities at the system-, science- and service-level in addition to the conventional large-capacity supercomputing approaches. The latest distributed architectures built around the composability of data-centric applications led to the emergence of an ecosystem for container coordination and integration. New approaches for dynamic composability of heterogeneous systems are needed to further advance the data-driven and AI-integrated scientific practice by multidisciplinary teams of researchers. This talk presents a novel approach for using composable systems for team science at the intersection of scientific computing, artificial intelligence (AI), and remote sensing at the edge, including the first working example of a composable infrastructure that federates Expanse, an NSF-funded supercomputer, with Nautilus, a Kubernetes-based GPU geo-distributed cluster and Sage, a reconfigurable edge AI infrastructure. It will also overview scientific workflow case studies that compose the insights from edge sensing, scientific instrumentation, AI, computing capabilities and physics-driven simulations in the wildland fire domain.

For decades, the computing continuum was simple — large scientific instruments and distributed sensor networks produced data, and advanced networking moved the data into data center repositories. “Data Movers” and networks built around a “Science DMZ” architecture were commonplace. However, bucolic computing in the HPC Shire has been permanently disrupted. Computation is no longer satisfied living in the data center, but wanders off to find adventure. Edge computing has upended traditional computation models — artificial intelligence can be run on the tops of mountains, on street corners in Chicago, and in the frozen land of Utqiagvik. From the Gold Coast of the West Island to the lands of the Ojibwe, edge computing and artificial intelligence has brought new adventures to the computing continuum. Climate, atmospheric, and environmental science are reimagining distributed sensing with AI-driven edge computing. What are the most monumental challenges for AI@Edge computation? Come and join the adventure to find out.

Did humankind’s confrontation with the coronavirus disease in late 2019 impact progress on delivering Commercial Off The Shelf computers that will process 10^18 floating operations per second?