China's LineShine Hits 2.198 Exaflops to Reclaim World's Fastest Supercomputer — Zero NVIDIA GPUs
Revealed at the TOP500 list in Hamburg on June 24, China's LineShine supercomputer beat the US El Capitan by 20% using entirely domestic CPUs — no NVIDIA or AMD hardware — demonstrating that US export controls failed to contain China's frontier compute ambitions.
China’s LineShine supercomputer posted 2.198 exaflops on the TOP500 list announced June 24 in Hamburg — a 20% lead over the United States’ El Capitan system at Lawrence Livermore National Laboratory, which holds 1.809 exaflops. This is China’s first time atop the TOP500 since 2017, ending nearly a decade of US dominance.
The specification that matters most is what LineShine does not contain: not a single NVIDIA or AMD GPU. The entire system runs on a custom domestic processor — a 304-core chip running at 1.55 GHz, with 13.79 million total cores across the system. It is the first machine ever to break 2 exaflops using a CPU-only architecture. No H100. No H200. No Blackwell. None of it.
That is the geopolitical headline inside the benchmark headline. US export controls on advanced chips — introduced in 2022 and tightened repeatedly since — were premised partly on the idea that restricting access to NVIDIA’s most powerful accelerators would slow China’s ability to build frontier compute infrastructure. LineShine is the counterargument in hardware form. China built the world’s fastest supercomputer by routing entirely around the restricted hardware.
The practical implications split across two domains. For scientific computing, LineShine gives China unprecedented domestic capacity for climate modeling, materials science, nuclear simulation, and AI training at national scale — without supply chain dependency on US export policy. For the chip industry, it validates years of investment in domestic CPU architecture that skeptics dismissed as too far behind NVIDIA’s accelerators to matter at the frontier.
The comparison to El Capitan is instructive. El Capitan is built on AMD EPYC CPUs and AMD Instinct MI300A accelerators — itself a deliberate choice to use non-NVIDIA hardware for energy and cost reasons. It reached 1.809 exaflops. LineShine surpassed that with a CPU-only approach on a custom architecture that did not exist as a commercial product five years ago.
The National Supercomputing Centre in Shenzhen operates LineShine. No official paper has been published yet, so the 304-core figure and 13.79 million core count come from the TOP500 submission data disclosed at the Hamburg event. Independent verification is pending — typical for new list entries — but TOP500 submissions require reproducible benchmark results.
This will have consequences for US export control policy. The argument that chip restrictions meaningfully cap China’s AI and HPC capacity becomes harder to sustain when China just demonstrated it doesn’t need the restricted chips to set a global performance record.