Le supercalculateur américain Frontier a pour la première fois de l'histoire passé le barre symbolique de 1 exaflop en juin 2022.

The world's most powerful supercomputer coming soon? Elon Musk's optimistic bet with Dojo

Article originally published on the "L'Usine nouvelle" website here

Tesla’s “Dojo” supercomputer, construction of which began this summer, is introduced as the world’s most powerful future supercomputer… by far. Elon Musk has announced that, with this cutting-edge equipment, he will be able to reach 100 exaflops before the end of 2024, thanks to a billion-dollar investment. An ambitious gamble: today, the most advanced computer is 100 times less powerful.

The American supercomputer Frontier passed the symbolic 1 exaflop mark for the first time in history in June 2022.

Invest twice as much to do a hundred times better? That’s Elon Musk’s challenge with “Dojo”. This supercomputer, whose construction announced by Tesla on July 19, will benefit from an investment of $1 billion over three years. The ambition: to create the world’s most powerful computer. And by far. With 100 exaflops expected for October 2024, the computer will be a 100 times more powerful than the world’s most powerful computer to date. It is destined to drive artificial intelligence models behind self-driving cars.

Surpassed for the first time in 2022 by the American computer Frontier, the exaflop barrier corresponds to the execution of 1 billion of billion floating-point operations per second. Germany and France are set to join the coveted club of exaflop supercomputers, in 2024 and 2025 respectively, with the European supercomputers Jupiter and Jules Verne. Production of the American computer began this summer, with the aim of entering the world’s top 5 at over 0.2 exaflops by January 2024. But the technical reality promises to be more complex.

A tight deadline

“Reaching 100 exaflops by the end of 2024 seems ambitious”, says Jean-Yves Berthou, Director of the Centre Inria in Saclay, France. It took Frontier almost a year to pass the exaflop mark. “Here, we would go from a situation with no machines to 100 exaflops in one year. It’s not impossible, but it’s very optimistic”, agrees François Bodin, professor at the Université de Rennes. The scientist highlights, for example, to the uncertainties surrounding the supply of GPU graphics chips, pantagrual data storage or the supply of electricity. Indeed, the average consumption of a supercomputer would be 20 megawatts, according to the CEA.

Tesla’s announcement also keeps the nature of these 100 exaflops unclear. The company has not specified whether this is a theoretical or actual capacity. The difference can be significant: corrected for factors such as memory latency or inter-node communication, Frontier achieves an actual capacity of 1.1 exaflops for a theoretical peak of 1.7 exaflops. To evaluate computer performance under real-life conditions, the supercomputers run the “Linpack” a test bench that brings together several programs and software libraries. “It’s not the same thing to 100 exaflops by running the Linpack algorithm producing 100 exaflops by running theLinpack algorithm than having a machine with a theoretical peak power producing 100 exaflops by running the Linpack algorithm than having a machine with a theoretical peak power of 100 exaflops”, sums up François Bodin.

A supercomputer like no other

Nevertheless, the major strength of Dojo’s computing power lies in its specialization. Frontier, Jupiter or Jules Verne have a wide range of applications – research in the fields of space, climate, pharmaceuticals, energy and many more. Dojo, on the other hand, has just one: training artificial intelligence models for autonomous cars. This type of calculation requires less precision than scientific calculations.

This difference enables Tesla to build a specialized processor architecture, with a lower number of bits (and therefore a lower precision of calculation). precision). “In all likelihood, they will run on 8-bit computation, rather than 64-bit as in classical scientific computing”, explains Jean-Yves Berthou. They will therefore gain a factor of 8 in power.” In other words, by agreeing to divide its calculation precision by 8, Dojo will be able to carry out 8 times as many instructions. We shouldn’t think that Europe is lagging behind, because it would invest half as much for a hundred times less performance”, argues the scientist. The exaflop metric is simply not used in the same context here.

Tesla’s gamble has already had a significant impact: the rise in its stock market price. A Morgan Stanley report estimates that Tesla’s market capitalization could rise by almost $600 billion (562 billion euros) thanks to the potential of its supercomputer, which is expected to help create robot cabs. As is often the case, Elon Musk’s announcements seduce investors. But beware : the billionaire already said he was “very confident” about the appearance of robot cabs in… 2020. His ambitions for Dojo could be just as over-optimistic.