China has taken the top spot on the TOP500 list of the world's fastest supercomputers with a system called LineShine, ending the run of El Capitan at Lawrence Livermore National Laboratory and marking the first time a China-based machine has led the rankings since Sunway TaihuLight in 2017. Reuters reports that the win comes with an important qualification: LineShine is built for traditional scientific computing, and the experts it interviewed do not read the result as China holding the fastest machine for AI work. The system carries no advanced AI accelerators, likely because the tools to make those chips remain subject to US export controls.
LineShine, installed at the National Supercomputing Centre in Shenzhen and built by the Shenzhen Cloud Computing Center, posted 2.198 exaflops on the High Performance Linpack benchmark, roughly 80 percent of its 2.736-exaflop theoretical peak. That makes it the first system in the rankings to sustain more than two exaflops of double-precision performance using CPUs alone. El Capitan now sits second at 1.809 exaflops, with Frontier at 1.353, Aurora at 1.012, and Europe's JUPITER Booster at 1.000.
The architecture explains the caveat. LineShine runs 13.79 million cores across 304-core LX2 processors at 1.55 GHz, tied together by a proprietary LingQi interconnect on the domestic LingKun platform and the Kylin operating system. There are no Nvidia or AMD GPUs in the machine. That all-CPU design is well suited to the double-precision scientific workloads the TOP500 measures, the simulations of atoms, climate, and fluid dynamics that have always defined high-performance computing. It topped the HPCG benchmark as well, at 22.00 HPCG petaflops.
On the benchmark that more closely resembles AI work, the mixed-precision HPL-MxP test, LineShine placed fourth at 7.92 exaflops. The ranking order flips because AI training rewards the dense, low-precision math that GPUs deliver, and an all-CPU system optimized for double-precision science does not lead there. "Fastest in the world" and "fastest for AI" have become two different titles, and LineShine holds the first without claiming the second. For anyone tracking the cooling implications, that distinction shapes the thermal design of the room.
LineShine draws approximately 42.2 megawatts and rates 52.07 gigaflops per watt. Nearly all of that 42.2 MW converts to heat that has to leave the building, and a machine packing 13.79 million cores does so at extreme density. A system at the top of the TOP500 implies a cooling plant operating at a correspondingly advanced level, because the compute record cannot be set without first solving the thermal one. The CPU-only design changes the shape of that load relative to a GPU machine, with heat spread across a very large processor count rather than concentrated in a few thousand accelerators, but the aggregate the cooling system must reject is the same order as any exascale peer.
China has been moving aggressively on the cooling technologies that make this density manageable. Domestic programs span cold plates, immersion, and materials work such as the diamond-copper composite coating developed by the Chinese Academy of Sciences to pull heat off silicon faster. The same export controls that kept advanced AI chips out of LineShine are also pushing China's domestic accelerator roadmap and the cooling demand that travels with it, and they help explain why an all-CPU exascale machine was the path of least resistance to the number-one slot.
The takeaway for the cooling industry is that the compute race is also a cooling-technology race, and the two now move together. China's ability to stand up a fastest-in-world system rests on a thermal plant capable of moving tens of megawatts at exascale density, and that capability is a strategic input rather than a back-office detail. The same forces driving China's more exotic compute ambitions run through cooling first, because every flop on the LINPACK score sheet is a watt that a heat-rejection system had to carry before the benchmark could finish.