China bypasses US GPU bans with 1.54-exaflops 'LineShine' supercomputer

In response to U.S. GPU export restrictions, China has successfully deployed a new CPU-only supercomputer named 'LineShine,' achieving a performance of 1.54 ExaFLOPS in BF16 training. The system consi... Read original →

China’s CPU-only 'LineShine' supercomputer represents a strategic pivot in HPC architecture, bypassing U.S. GPU bans through Armv9-based LX2 processors with integrated SVE/SME and hybrid HBM/DDR5 memory. While sacrificing peak AI throughput, it gains in scientific workloads requiring low-latency communication, accelerating domestic EDA, interconnect (LQLink), and Arm software stack development. Compliance-wise, reliance on foreign accelerators drops, but 3nm scaling without EUV access risks cost and yield. NVIDIA may counter by tightening Grace-Hopper integration, while Huawei pushes Ascend adoption. Within 18 months, global HPC will bifurcate: heterogeneous acceleration dominates the West, while China advances CPU-native AI, elevating mixed-precision FP64/BF16 as the new norm for AI-for-science and intensifying the battle for chip architecture sovereignty.