At the International Supercomputing Conference in Hamburg, organizers announced that LineShine, the new flagship system at Shenzhen’s National Supercomputing Center, had topped the TOP500 list with a benchmark score of 2.198 exaflops. The figure places the machine more than 20 percent ahead of the previous leader, the U.S. Department of Energy’s El Capitan, and ends a long run of American dominance on the most closely watched supercomputing ranking.

What makes LineShine’s triumph remarkable is not just the raw speed but the way it was built. Unlike most top‑tier machines that rely on graphics processing units from Nvidia or AMD, LineShine runs solely on custom CPUs built on the Armv9 architecture. The processors, known as LX2, were developed in China and are paired with a home‑grown network called LingQi and the KylinOS operating system, a Chinese fork of Linux. The design eliminates any reliance on U.S. chips, a direct response to export controls that have limited American semiconductor sales to Chinese firms for years.

The all‑CPU architecture packs nearly 14 million cores into 90 cabinets and consumes roughly 42 megawatts of power, delivering about 52 gigaflops per watt—a notable efficiency for a system of its size. Already, LineShine has run a full Earth‑system simulation and a detailed model of the human brain, showcasing its capability for the heavy scientific workloads it was built to handle.

Jack Dongarra, a senior scientist with the TOP500 project, visited the facility and praised the machine’s performance. “They upped us by developing a system that is not reliant on GPUs,” he said, highlighting the strategic shift away from the graphics‑centric designs that dominate the field.

China had stopped submitting entries to the TOP500 in 2023 after the United States tightened chip‑export rules. By entering LineShine, the Chinese team sent a clear signal: the nation can achieve world‑class performance without foreign technology. Analysts note that the move also exposes a regulatory gap—while GPUs are tightly controlled, CPUs face far looser restrictions, allowing China to advance its own high‑performance computing stack.

Despite its record‑breaking score on the standard benchmark, LineShine trails in AI‑focused tests. On a mixed‑precision benchmark that mirrors artificial‑intelligence workloads, the system placed fourth with 7.92 exaflops, far behind GPU‑heavy rivals. The all‑CPU design lacks the low‑precision circuitry that gives GPUs a speed edge in AI model training. Consequently, while LineShine holds the title of the world’s fastest computer for double‑precision scientific calculations, it does not yet lead the AI race.

The achievement underscores a broader trend: U.S. export controls intended to slow China’s tech progress have spurred the development of a parallel domestic stack of chips, operating systems, and interconnects. Companies across sectors—from automakers to cloud providers—are now designing custom silicon to reduce dependence on foreign GPUs. Even Nvidia is exploring Arm‑based CPUs to stay competitive.

Experts caution that the crown does not signify a complete closing of the gap with the United States. China’s domestic GPU industry still lags behind Nvidia and AMD, and the most powerful AI clusters run by American hyperscalers remain outside the TOP500’s academic focus. Nonetheless, LineShine’s success demonstrates that a policy aimed at curbing China’s access to advanced chips may be prompting the country to build its own high‑performance computing ecosystem.

Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.