The Lingxing LX2 includes integrated high-bandwidth memory (HBM) within the package, showcasing advanced packaging amid export controls on memory and components.

LineShine, China's CPU-focused supercomputer, has overtaken El Capitan to lead the Top500 list with a Linpack FP64 performance of 2.198 exaflops, marking a historic CPU-first exascale milestone.

The system runs on standard CPUs and Arm-based LX2 chips, not GPUs, and achieves about 42.2 megawatts of power with an efficiency around 52.07 gigaflops per watt.

LineShine’s architecture integrates GPU-like acceleration directly into its CPUs and features nearly 14 million cores across 90 cabinets, using an embedded accelerator circuitry to speed up matrix and vector work without NVIDIA GPUs.

Analysts point to ongoing concerns about energy and water use in data centers and potential climate impacts, even as investments in AI and HPC continue.

The LX2 processor lineup relies on Arm architecture, with no disclosure of specific manufacturers or production partners for LineShine’s CPUs.

LineShine’s CPU-first design is notable for diverging from the GPU-centric path of most top systems and is viewed as a potential model for combining AI with traditional scientific computing.

Reuters notes the result may reflect Beijing’s goal of signaling self-sufficiency in computing rather than declaring AI leadership, given evolving industry practices and list methodologies.

El Capitan, operated by Lawrence Livermore National Laboratory, ranks second, with other Top10 entries including systems from Tennessee and Illinois in the U.S. and Germany’s Jupiter in the fifth spot.

China’s policy push for domestic tech and reduced reliance on foreign hardware frames LineShine within a broader tech competition, though the country’s GPU industry remains behind leading GPUs.

The discussion contrasts traditional high-precision 64-bit HPC workloads with AI-oriented, lower-precision computing, highlighting differing optimization goals.

LineShine’s compute nodes contain 152 cores per die and 128 GB of off-package memory across four NUMA domains, with a dedicated SDMA engine for DDR-HBM data movement.