LineShine Tops TOP500 Without Nvidia, Intel Or AMD Chips

LineShine Takes The TOP500 Lead With Local CPUs

China's LineShine supercomputer has taken first place on the TOP500 list with 2.198 Exaflop/s of sustained double-precision performance, doing so without Nvidia, Intel or AMD processors.

The system is housed at the National Supercomputer Center in Shenzhen and uses LX2 processors described as a local effort.

The technical detail is more complicated than a clean break from foreign technology because the processors use Armv9 designs and the machine runs KylinOS, a Linux distribution.

LineShine comprises 20,480 computing nodes.

Each LX2 processor integrates two compute dies with 304 cores in total and eight on-package HBM stacks with 32 GB and 4 TB/s of aggregate bandwidth.

Each compute die carries 152 cores.

The design pairs off-package DDR memory with four NUMA domains, and a dedicated SDMA engine moves data between DDR and HBM.

The pre-press paper described LX2 support for FP64, FP32, FP16 and INT8 through SME and SVE units.

It puts per-processor performance at up to 60.3 TFLOPS in FP64 and 120.6 TFLOPS in FP32, adding another technical marker to the system's CPU-only claim.

CPU-Only Exascale Creates A Different AI Compute Signal

LineShine became the first TOP500 system to exceed two exaflops of sustained double-precision performance using CPUs only.

TOP500 curators said the system could perform better in future tests because it reached about 80 percent of its 2.736 Exaflop/s theoretical peak.

The interconnect is part of the sovereignty claim.

Nodes are linked through the LingQi high-speed network, a dual-plane multi-rail fat-tree topology that offers 1.6 Tb/s bandwidth per node.

The network comes from Hangzhou LingQi Technology Co.

The result gives Beijing a concrete example of domestic high-performance computing at a time when the government is steering local organizations toward made-in-China technology.

It also gives Chinese AI and scientific computing programs a flagship system at the top of the global ranking.

The ranking does not put China in control of the whole list.

Nvidia, AMD and Intel still dominate this TOP500 edition, even as LineShine gives China the leading individual system.

The technical stack still shows dependence on international layers.

Arm architecture and Linux keep LineShine tied to technologies and communities outside China, even as the processors and interconnect are presented as local components.

Export Controls Keep The Supply Chain Question Open

The ranking arrives while US restrictions limit GPU sales to China.

Those controls have made compute supply chains a direct policy issue for AI, military modernization and industrial planning.

LineShine does not remove that pressure.

The machine demonstrates that China can assemble a world-leading CPU-only supercomputer, but the article does not establish whether the design can replace restricted GPUs for the full range of AI workloads.

China's GPU industry remains early, with current products described as trailing Nvidia and AMD by four or five years.

That keeps LineShine's CPU result separate from the accelerator supply problem created by export controls.

The system's benchmark also leaves scaling questions beyond the TOP500 result.

Sustained double-precision performance is central to high-performance computing rankings, while AI training and inference depend on software, memory movement, accelerators, networking and workload-specific economics.

LineShine now gives China the leading TOP500 machine, 20,480 nodes and a local interconnect, while its Armv9 base, Linux operating system and unresolved GPU-substitution question keep the compute-sovereignty story tied to global technology dependencies.