LineShine, a 2.198 exaflop supercomputer, tops TOP500 without Nvidia, Intel, or AMD

The TOP500 list just crowned a new king, and the headline number matters: LineShine delivers 2.198 Exaflop/s, topping the world’s most powerful supercomputers. The twist is how it got there. The system reached first place without using any kit from Nvidia, Intel, or AMD, which is exactly the kind of procurement and ecosystem assumption that usually goes unchallenged in this space.

LineShine is housed at the National Supercomputer Center (NSC) in Shenzhen, and the underlying details explain why it is more than a symbolic win. In pre-press work referenced by the report, the machine’s LX2 processors are described as a local effort that uses Armv9 designs, keeping Beijing’s “indigenous everything” story tightly connected to the global chip ecosystem rather than fully breaking away. LineShine also runs KylinOS, a Linux distribution with contributions from around the world, and the curators of TOP500 think the system could do even better in future tests. This time, it reached about 80 percent of its 2.736 Exaflop/s theoretical peak in preparation tests for the current TOP500 iteration.

So what exactly is inside this thing? The paper reveals LineShine comprises 20,480 computing nodes. Each LX2 processor integrates two compute dies with 304 total cores, plus eight on-package HBM stacks providing 32 GB per processor and 4 TB/s aggregate bandwidth. Each compute die has 152 cores and 128 GB of off-package DDR memory organized into four NUMA domains. A dedicated SDMA engine handles data movement between DDR and HBM, which matters because supercomputers do not fail only on raw compute, they fail when the data plumbing cannot keep up.

Compute capability is only half the story, and the report leans into the other half: the interconnect. Nodes are connected via the LingQi high-speed network, using a dual-plane multi-rail fat-tree topology. That network is also described as a Chinese creation, provided by Hangzhou LingQi Technology Co. Per node bandwidth is stated as 1.6 Tb/s. In practical terms, that is the difference between a cluster that can crunch and a cluster that can scale, especially for workloads that require constant synchronization.

The performance milestone the report highlights is unusually specific: LineShine became the first system on TOP500 to exceed two exaflops of sustained double-precision performance using CPUs only. The LX2 platform supports FP64, FP32, FP16, and INT8, via SME and SVE units, with stated up-to 60.3 TFLOPS in FP64 and 120.6 TFLOPS in FP32. That CPU-only “sustained double precision” phrasing is not marketing fluff. It signals that this approach is not just optimized for a narrow benchmark, it is aiming at the kind of numerical accuracy supercomputing historically cares about.

Why does this matter beyond a bragging-rights scoreboard? Because the timing is political and economic in a very direct way. The report frames LineShine’s rise as coming when China’s government increasingly steers local organizations toward buying made-in-China technology. The strategic intent is to decrease dependency on foreign products as Beijing pursues AI and other technologies to boost economic growth and enhance military capabilities. In that storyline, import reliance can stymie those ambitions, and the report points to the USA’s ban on GPU sales to the Middle Kingdom as evidence that supply chain control is now a national policy lever.

There is another layer executives should notice: the message to boards and procurement teams is that “the stack” may be reconfiguring under their feet. The report notes that Nvidia, Intel, and AMD dominate this version of TOP500, and that China is still on a long march toward global dominance. It also claims China’s GPU industry is nascent, with products trailing Nvidia and AMD by four or five years. Read together, those two points imply a dual strategy. Keep building CPU-centric high performance now, while GPU ecosystems mature later. That can reduce near-term bottlenecks while aligning with industrial policy.

For leaders in similar roles, the strategic stakes are straightforward. If a CPU-only approach can hit 2 exaflops sustained double precision and claim TOP500 leadership without three dominant suppliers, then future planning for compute procurement, data center architecture, and vendor risk can no longer assume the old default hierarchy. Even if China is not “winning everything” yet, it is demonstrating credible capability in the exact category that used to belong almost entirely to the established Western supply chain. That is the kind of development that forces every operator to ask harder questions about resilience, supply continuity, and what “performance” really means when the interconnect, memory bandwidth, and operating system ecosystem all become part of the competitive equation.