Exascale Computing

In the simplest and most literal sense, exascale computing refers to computing systems capable of a certain speed/throughput threshold—specifically 10^18 FLOPS. But in a broader sense, according to Youssef Marzouk, professor in the Dept. of Aeronautics and Astronautics at MIT and codirector of MIT’s Center for Computational Science and Engineering, it is a threshold that embodies a new frontier of computing at scale. The new frontier holds promise, but not without challenges.

“Building computing hardware that can, in principle, achieve this threshold is only part of the challenge,” Marzouk says. “The other part is to create software that can actually use this hardware effectively. My main interest is to use exascale computing to simulate complex physical systems with the goal of solving open problems in science and engineering. Here, the challenges are manifold. Being able to write efficient programs for proposed exascale computing systems is a challenge unto itself; new programming languages and parallel compiler technologies are needed to make this possible.”

Another piece of the exascale puzzle is what goes in the codes themselves. “In a physical setting, we can develop simulations with billions of atoms or billions of grid cells, but how do we know that they are actually predictive?” asks Marzouk. “How can we test or ensure that they represent our physical reality, at least within some bounds or sense of uncertainty? There are so many interesting issues underlying this question at the intersection of mathematics, statistics, computational science, and physical modeling.

Berzins and many others theorize that the first exascale machine will likely be the DOE Oak Ridge National Laboratory’s Frontier sometime this year. “The Intel Aurora architecture at the U.S. DOE Argonne National Laboratory is likely to follow in 2022,” Berzins adds. “However, there are also three possible exascale machines being developed in China—one or more of which may come online in the same timeframe.”

Doug Kothe, director of the ECP (Exascale Computing Project) at Oak Ridge National Laboratory, says capable (e.g., affordable, usable, and useful) exascale computing will be in the U.S. within two years. But he also says achieving exascale is about more than reaching 10^18 FLOPS. “For the U.S. DOE, it is a capable computing ecosystem (applications, software stack, hardware) specifically codesigned to provide breakthrough solutions addressing our nation’s most critical challenges in scientific discovery, energy assurance, economic competitiveness, and national security,” Kothe explains. “This ecosystem is not just a matter of ensuring and relying upon more powerful computing systems, but rather it must foster and support more valuable and rapid insights from a wide variety of applications, which requires a much higher level of inherent efficacy in all methods, software tools, and exascale-enabled computing technologies.”

AI and ML (machine learning) are becoming increasingly important focal points within supercomputing. “Over the last 30 years, the predominant use of supercomputers has been to support the most challenging modeling and simulation applications,” Parsons says. “There has been little focus on AI and machine-learning technologies. However, over the past few years, as data science challenges have grown across the commercial and academic sectors, the use of supercomputers to process large amounts of data has grown. Couple to this (the fact that) it is now commonplace for the largest supercomputers to contain large numbers of accelerators based on GPUs. This makes these systems ideal platforms for the largest AI and machine-learning problems. At the same time, the features of GPUs that make them very good at the training step of deep learning are gradually being added to CPUs (low and mixed-precision floating point arithmetic and matrix operations, for example). This means that exascale systems, which—with one honorable exception—are based on a node design with a CPU plus multiple GPUs will be the perfect platform for the largest AI training challenges.”

Simon David Hammond, research scientist in the Scalable Computer Architectures group at Sandia National Laboratories, says exascale’s application in the AI and ML realms is an interesting area of discussion in the industry today. “Exascale computing is pushing the limits of computational power within specific energy, power, and cost targets,” Hammond says. “By redesigning high-performance systems to meet exascale targets, the entire computing industry is much more aggressively trying to find new designs to meet these targets, and, as they do, many of these techniques will be equally useful within the AI and machine-learning technology space.”

The question is also whether ML will be used within the scientific domain to support activities that today are thought of as HPC—for instance, replacing high-performance computing simulations with machine-learning interference.”

Hammond says exascale-class systems will impact a huge swathe of science and engineering industries, and that’s what makes exascale so important. “Some of the most exciting areas will be in much more capable engineering design capabilities; for instance, designing much higher efficiency and cleaner engines, more efficient planes, cars, and trucks, and novel approaches to energy generation, such as efficient wind turbine designs for renewables,” he explains. “By using virtual design capabilities, we should be able to optimize designs and do this faster and at lower cost than ever before. Another area which has the potential to be truly transformative is in medicine, either through the development of more efficient drugs or through the creation of drugs which more closely map to an individual’s personal genetics. I could imagine a world where virtual drug design and evaluation helps to reduce development costs and improve medical outcomes.”

These 24 mission-critical exascale applications include the predictive microstructural evolution of novel chemicals and materials for energy applications; the ability to accelerate the widespread adoption of additive manufacturing by enabling the routine fabrication of qualifiable metal alloy parts; the ability to demonstrate commercial-scale transformation energy technologies that curb fossil fuel plant CO₂ emission; the ability to address fundamental science questions; the ability to reduce major uncertainties in earthquake hazard and risk assessments to ensure the safest and most cost-effective seismic designs; the ability to forecast water resource availability, food supply changes, and severe weather probabilities with confidence; the ability to optimize power grid planning and secure operation with very high reliability; and the ability to develop treatment strategies and pre-clinical cancer drug response models and mechanisms for certain cancers.

Simon McIntosh-Smith, professor of high-performance computing and head of the HPC Research Group in the Dept. of Computer Science at the University of Bristol, says exascale computing will enable breakthroughs in a wide range of application areas, but there are four in particular that excite him the most. First, there’s weather and climate modeling. McIntosh-Smith says the resolution and sophistication of the models will increase dramatically, enabling hyperlocal weather predictions, better forecasting of dangerous weather extremes like storms, floods, snow, and heat waves, and a better understanding of humanity’s impact on our climate. Second, there’s rational drug design. Exascale may allow us to optimize drugs and vaccines, designing and optimizing them for a very specific need.