China's LineShine tops TOP500 with CPU-only design

The design puts scalar, vector, matrix, control-flow, communication and I/O operations into a single CPU execution environment, so simulation and AI stages share one memory hierarchy and one software stack.

LineShine leads on double-precision HPL and on HPCG but ranks fourth on HPL-MxP, the mixed-precision test closer to AI training and inference. TechRepublic noted that a system can lead the main list and still trail on the benchmark more closely tied to AI infrastructure.

What LineShine actually won

LineShine sits at the National Supercomputing Center in Shenzhen and recorded 2.1984 EFlops of sustained HPL performance, the first CPU-only system on the TOP500 to pass 2 EFlops. El Capitan at Lawrence Livermore National Laboratory, the previous titleholder, is listed at 1.809 exaflops. Global Times reported that LineShine was approximately 20 percent faster than El Capitan.

The rest of the top five was familiar: El Capitan second, then Frontier and Aurora in the US and JUPITER in Germany. Exascale machines are now spread across the US, Europe and China, though the US still holds several of the highest-ranked systems.

At full scale the machine runs about 90 compute cabinets and 22,680 compute nodes, more than 13 million processor cores in all. Each LX2 CPU carries 304 cores, and each node holds two of them, alongside 64 GB of HBM and 512 GB of DDR5 memory. It draws about 42.2 megawatts.

LineShine also took first place on HPCG with 22.0049 PFlops, a benchmark weighted toward memory access, communication and sparse computation rather than dense arithmetic.

The AI benchmark tells a different story

On HPL-MxP, which tests mixed-precision performance, LineShine ranked fourth with 7.92 exaflops, while GPU-accelerated systems did better on the lower-precision operations common in AI training and inference. TechRepublic noted that HPL measures sustained double-precision performance, while HPL-MxP is closer to the arithmetic used in many AI workloads.

The data released with the results shows LineShine does not include advanced AI chips.

Addison Snell, quoted by mezha.net, said that if the hyperscalers had submitted their systems, this "world's fastest" would not have made the top five. He said he was not surprised the system took first place, but was surprised it was submitted and that recognition was sought for it.

Nine years, and a decision to submit

China last held the top spot with Sunway TaihuLight in 2017, a gap Global Times and Newswise both count as nine years. China stopped submitting systems for ranking in 2023, after years of US export controls on chips and computing technology.

The Global Times editorial traced the controls back to 2015, when the US Commerce Department listed four Chinese supercomputing centers associated with Tianhe-2. It then cited 2019, when Sugon and Sunway TaihuLight were targeted, and seven more entities added in 2021.

The New York Times quoted Jack Dongarra, an organizer of the TOP500 list, as saying LineShine is "an impressive system." He added that China had upped the US by developing a system not reliant on GPUs, Global Times reported. Zhang Yunquan, a researcher at the Institute of Computing Technology under the Chinese Academy of Sciences, told the Global Times that LineShine follows a pure CPU route on the ARM architecture, unlike most US systems that pair CPUs with GPUs.

Yutong Lu, the system's chief designer, said in the Frontiers of Computer Science paper that future scientific computing will run simulation, AI training and inference, optimization and data processing within one continuous workflow.