China Deploys 1.54-ExaFLOPS LineShine With 2.4 Million Armv9 Cores, Bypassing US GPU Bans
Updated
Updated · Tom's Hardware · May 17
China Deploys 1.54-ExaFLOPS LineShine With 2.4 Million Armv9 Cores, Bypassing US GPU Bans
1 articles · Updated · Tom's Hardware · May 17
20,480 nodes at China’s National Supercomputing Center in Shenzhen now power the CPU-only LineShine system, which uses 40,960 LX2 processors and 2,451,840 Armv9 cores for AI and HPC workloads.
1.54 exaFLOPS of BF16 training performance comes from custom LX2 chips built around matrix and vector extensions, with each processor pairing 32 GB of HBM and up to 256 GB of DDR5.
2.16 exaFLOPS was the peak reached while training a 6.3-billion-parameter Earth-observation generative compression model, while theoretical FP64 peak performance is 2.47 exaFLOPS.
US restrictions on advanced GPUs have pushed China toward CPU-only supercomputers that avoid dependence on Nvidia-style accelerators and CUDA, though the tradeoff is lower dense-AI efficiency and weaker power efficiency than GPU-heavy systems.
Can China's homegrown supercomputer chip truly challenge Nvidia's dominance in the global AI race?
Did US tech sanctions just accelerate China's rise as a self-reliant supercomputing power?
In 2026, China unveiled LineShine, a fully domestic supercomputer that marks a strategic shift towards CPU-only designs for exascale computing. This breakthrough system is built in two phases, each adding to its massive processing power and data capacity. LineShine is designed to handle complex scientific and AI applications that often struggle on GPU-centric systems, showcasing advanced capabilities and reinforcing China’s self-reliance in critical computing infrastructure. Its development represents a major technological achievement and signals a new era for domestic high-performance computing.