Frontier vs. Aurora: The Battle for the Title of America’s Fastest Supercomputer

The first thing you notice when you enter Frontier’s wing of Oak Ridge National Laboratory is not its size. It’s the noise. A long, low hum that resembles the weather. For years, the fastest computer on the planet was responsible for that hum, and those who worked close to it spoke about it with a peculiar blend of pride and superstition, much like pilots talk about their favorite aircraft.

After years of delays that had turned into a running joke in the world of high-performance computing, Aurora finally woke up in Illinois. No supercomputer in recent memory may have been written off as frequently or too soon. Argonne engineers insisted that it would arrive. Around 2022, the majority of the field stopped paying attention.

This contributes to the current rivalry’s sense of intimacy. In the summer of 2022, Frontier became the first company to cross the exascale threshold, and it remained at the top for five consecutive TOP500 rankings. a lengthy reign in a field where everything can change in just six months. Similar to a sprinter who shows up to the final with one shoe untied, Aurora arrived second, half-finished, and scored respectably but obviously incomplete.

The numbers now reveal a more subdued narrative. At 1.353 exaflops, Frontier is positioned. At 1.012, Aurora has crossed the exascale threshold, which sounds slower until you keep in mind that the system isn’t finished. When Aurora is fully online, it is anticipated to surpass two exaflops, according to Argonne staff, who have made it clear that Aurora is still being tuned and brought up node by node. As this develops, it seems as though Frontier’s lead is genuine but temporary.

Raw speed isn’t what makes the comparison intriguing. Its purpose. Frontier was designed to be a general scientific workhorse, capable of performing tasks that warrant a federal budget line, such as materials science, fusion simulation, and climate modeling. Aurora was created with an alternative obsession in mind. It is adjusted for AI. It has reportedly reached 10.6 exaflops on AI workloads, compared to Frontier’s 2.35. Everyone in the room is aware of how quickly the science supporting these machines is changing, so that gap is more important than the headline figures.

The chip question is another. All of Frontier’s CPUs and GPUs are made by AMD, giving the company a narrative it had been waiting decades to share. Aurora took a different approach, using an entire stack from a company that hasn’t had a great few years in the data center, including Intel Xeon Max processors and Intel Data Center GPU Max accelerators. If Aurora fulfills its potential, it subtly restores a portion of Intel’s reputation that was previously dismissed by the majority of analysts. That is not insignificant.

Aurora finds the topic of energy efficiency to be uncomfortable. Frontier provides roughly 55 gigaflops per watt. As of right now, Aurora is closer to 26. That disparity is difficult to overlook in a time when data center power consumption is turning into a political issue as much as an engineering one. Whether tuning reduces the gap or if this is a result of Intel’s design decisions is still up for debate.

El Capitan at Livermore, which currently tops the list, eventually eclipsed both systems. However, the rivalry between the Frontier and Aurora is still the topic of conversation in the industry. Perhaps because it’s a tale of ambition that took longer than necessary and a champion who most likely knows that his time is running out. It’s difficult to avoid thinking that the next ranking will read very differently as you watch this develop.