“I don’t think anybody here is ignorant of what supercomputing is,” said Rev Lebaredian, Nvidia’s vice president of Omniverse and Simulation Technology, as he opened the first keynote at ISC 2022 in Hamburg, Germany. “We’ve been building supercomputers for decades, but our uses for them have been evolving over time.” In his keynote, Lebaredian made the case for what he views as the next necessary evolution of supercomputing: effective digital twins.

Supercomputing’s traditional uses, Lebaredian (pictured in the header) said, have been persistent for a long time—uses like fluid dynamics, climate analysis, seismic modeling and so forth. Somewhat more recently, fields like drug discovery and renewable energy development have made their way into the HPC mainstream, fueled by the increasing capabilities of AI and its integration with high-performance computing.

Next, he said, this supercomputer-powered AI should be combined with computer graphics to enable broader and deeper use of digital twins—virtual facsimiles of real-world environments that are functionally indistinguishable from their physical counterparts.

This is not a game

“Now, when I say computer graphics, probably most of the people here who are in the sciences or in industry, you think of it as sort of a fanciful thing—we create photorealistic images primarily for entertainment, whether it’s visual effects, or video games—something like that,” Lebaredian said. “Which is all true—that’s all cool stuff. I came from visual effects, I’ve worked on video games while I’ve been here at Nvidia for 20 years or so.”

“But,” he continued, “if you look at the origins of computer graphics, 3D computer graphics in particular, the purpose is to create images but we do that by essentially simulating a virtual world. We create facsimiles of the real world in three dimensions, try to understand the fundamental laws of physics, and eventually turn that into pixels or into some sort of image we can see.”

Lebaredian showed examples of how early graphics experiments often focused on replicating the real world, stressing that the same work was important now to create realistic virtual worlds.

“If you can construct a virtual world that matches the real world in its complexity, in its scale, in its precision, then there’s a great many things you can do with this,” he said. “You can use it as a playground for the artificial intelligences that we’re creating—this is where they can get trained, this is where they’re born and raised and they gain experience—we can use it to understand our world better, we can use it as a lab to do experiments well before we try those things in the real world.”

“We need to reconstruct exactly how the factory looks for the AIs. They’re going to learn how to see, and if they learn how to see in a cartoon world, it’s not going to work in the real world. This photorealism isn’t about making pretty pictures, it’s about matching exactly what those sensors are going to experience so that the intelligence that we create is going to be correct.”

Lebaredian said that for these worlds to approximate reality, they need three things: rendering, or the interaction between light and matter; the general physics that define how that environment behaves; and simulated intelligence (i.e., AI). “Combining these three things—how the world looks, how the world behaves, and how things inside this world act upon the world—gives you all of the elements you need to create a virtual world. And to power it, you need a supercomputer.”

Gaining superpowers

Most products, buildings and so forth, Lebaredian said, are already created in the digital world through some form of computer-aided design (CAD) process.

“The first twin that’s born is usually the digital one, not the physical one,” he explained. “But that’s not where it ends. At this point in most cases they diverge: we forget about the original digital version and we continue to iterate and modify the physical version and they get out of sync and they cease to be twins. What we’re trying to introduce now is a mechanism by which we can link the two together, where we can detect all of the changes that are happening inside the physical version … through sensors, through the controllers that we use, and even the humans that are acting inside these worlds and reflect them in the digital world.”

“If we can establish that link,” he continued, “we gain some amazing superpowers.”

First among these superpowers: teleportation, the ability to use a digital twin to visit a factory, facility or planet that is far away from your real-world location. “Over time we’re going to make that feeling of teleporting there more immersive and more indistinguishable from the real world,” Lebaredian said.

Second: time travel. “If you record the state of the world over time and you have that persist in your storage, you can recall it at any point, and this allows us to time travel, essentially, into the past,” he said, explaining that this power could be used to, for instance, reconstruct a car crash to learn what caused it.

But this isn’t where the “superpowers” of digital twins stop. If you then incorporate a simulator that is accurate and predictive, Lebaredian said, you then have time travel to the future—and that allows for the testing of alternate futures, helping users to “find the best possible future” that they “would actually like to implement.”

What’s needed

“To do all of this we need a lot of new technologies—ones that don’t quite exist yet but [which] we’re on the cusp of actually realizing,” Lebaredian said. First up: computers with extremely precise time synchronization. “That’s the only way we’re going to be able to scale up this compute and do it fast enough and in real time with the latencies that we need to solve this problem.”

On the software side, digital twins will require new simulation technology—in particular, AI that is savvy enough to learn differences between the real world and simulations and automatically adapt the simulations to better reflect the real world.

As a result, the supercomputing that powers these simulations needs to be able to push AI capabilities to the maximum. “This means deviating from traditional supercomputing with double-precision, 64-bit calculations being the number-one thing to various types of calculations going down to low-bit precisions like 8-bit or even lower,” Lebaredian said. And: “To accelerate the algorithms that describe the rules of our world, we need all of the networking that will pump the immense amounts of data—the crazy amounts of data—that the real world has.” (All of this, of course, just so happens to be part of Nvidia’s value proposition.)

The computing will also need to extend to the edge in real-time. “If you’re controlling a robot, it can’t wait for the computation to happen on the other side of the Earth hundreds of milliseconds away before it gets its control signal back,” he said.

Bringing digital twins to the real world

Toward the end of his keynote, Lebaredian introduced Michele Melchiorre, BMW’s senior vice president for production systems, technical planning, tool shop and plant construction. At BMW, Melchiorre is working with Nvidia to implement a digital twin of its forthcoming factory in Hungary, which just broke ground this week. “This will be the first plant where we’ll have a complete digital twin much before the car production starts,” Melchiorre said.

This digital twin, he said—unlike the factory itself—is “already 80 percent finished.” When finished, Melchiorre said that the digital twin will help BMW optimize its workflows and make modifications before committing to real-world design choices, increasing speed and reducing cost. The digital twin will also streamline remote collaboration—a process that Melchiorre said was crucial in an era where it was sometimes almost impossible to meet together in person due to Covid restrictions.

Of course, hurdles remain for the digital twin. Just the body shop of the factory comprises some 20PB of data, and turning such massive amounts of data into functional, real-time simulations means engaging in some serious streamlining. “So not even for a simple die [do] we have enough computing power to do a perfect simulation,” Melchiorre said.

“The real world is glorious in its complexity,” Lebaredian added, “and there’s a certain amount of hubris we have here trying to recreate it in the virtual world.”