China’s LineShine Tops the World’s Supercomputer Rankings — Without a Single GPU

China’s LineShine Tops the World’s Supercomputer Rankings — Without a Single GPU

Built entirely from domestic chips, China’s new exascale machine at the National Supercomputing Centre in Shenzhen has dethroned the American El Capitan with 2.198 exaflops of sustained CPU-only performance — the first system ever to cross the two-exaflop threshold using processors alone.

A New Number One

The 67th edition of the TOP500 list, unveiled on June 23, 2026, at the ISC High Performance conference in Hamburg, Germany, delivered one of the most dramatic reshuffles in the ranking’s three-decade history. China’s LineShine supercomputer debuted directly at first place, displacing the United States’ El Capitan — which had itself only recently claimed the crown — and giving China its first number-one ranking since Sunway TaihuLight held the top spot back in 2017.

LineShine is installed at the National Supercomputing Centre in Shenzhen (NSCS) and was built by the Shenzhen Cloud Computing Center. Its defining characteristic is architectural: in an era when virtually every competing leadership-class system relies on GPU accelerators for raw throughput, LineShine achieves its record performance using CPUs alone. There is no Nvidia silicon inside, no AMD Instinct accelerator, no dedicated AI chip of any kind — only custom Chinese processors operating at scale.

The Hardware Inside

The system is built around the proprietary 304-core LX2 processor, clocked at 1.55 GHz and based on the Armv9 instruction-set architecture. Each chip incorporates Arm’s Scalable Vector Extension (SVE) and Scalable Matrix Extension (SME) units, enabling it to handle the vector and matrix operations central to both scientific computing and AI workloads across data formats including FP64, FP32, BF16, FP16, and INT8. Crucially, every LX2 pairs 32 GB of on-package HBM — a first for a Chinese-made CPU — with up to 256 GB of external DDR5 memory, delivering up to 4 TB/s of memory bandwidth per chip.

Across the full installation, LineShine fields approximately 13.79 million cores in total — an extraordinary number that underscores how the system achieves its record: through sheer parallelism rather than the per-chip efficiency advantages that GPU-based designs typically enjoy. The nodes are linked via a proprietary LingQi interconnect running at 1.6 Tb/s per node, using a four-layer fat-tree topology with an optical layer to sustain coherent communication at scale across up to 2 million ports and 100,000 nodes. The operating system is China’s own Kylin OS, and the entire system uses 100% liquid cooling.

Performance Across Benchmarks

On the standard HPL (High Performance Linpack) benchmark, LineShine recorded 2.198 exaflops — roughly 80 percent of its 2.736 exaflop theoretical peak, a utilization rate that actually compares favorably to the GPU-based American exascale machines. The system also claimed first place on the HPCG (High Performance Conjugate Gradient) ranking, posting 22 HPCG-PFLOPS/s, which better reflects real-world scientific application performance than Linpack alone.

The picture is more nuanced on the HPL-MxP mixed-precision benchmark, which measures the lower-precision math that AI training workloads actually depend upon. There, LineShine placed fourth with 7.92 exaflops — a multiplier of only 3.6 times its standard Linpack score. By contrast, El Capitan achieves a 9.2× uplift and Aurora a 11.5× uplift on the same benchmark. The comparatively modest mixed-precision gain is the expected signature of a system without dedicated low-precision accelerators, and reflects the deliberate design philosophy: a machine built for broad scientific computing rather than specifically for AI training throughput.

Power and Efficiency

LineShine draws approximately 42.2 megawatts at full load — an enormous appetite that reflects the cost of reaching the top through CPU scale rather than GPU efficiency. Its energy efficiency figure of around 52 GFLOPS/W trails El Capitan’s roughly 61 GFLOPS/W, meaning that while LineShine is the faster machine in absolute terms, it is less efficient per watt. The Green500 energy-efficiency ranking, which measures GFLOPS per watt, remains led by smaller European systems; LineShine does not challenge for that title.

“Export controls may slow China’s access to certain advanced components, but they also provide a strong incentive to develop domestic alternatives. LineShine suggests that China has responded through large-scale investment and hardware-software codesign.”

— Jack Dongarra, emeritus professor, University of Tennessee; co-organizer of the TOP500 list

A Full-Stack Domestic Achievement

Perhaps the most significant dimension of LineShine’s debut is not the benchmark number but what it is made of. Every major component — the LX2 processor, the LingQi interconnect, the storage subsystem, the Kylin operating system, and the full-stack software environment — is of Chinese origin. The system also represents the first Chinese-made machine to deploy HBM memory, boosting bandwidth by up to ten times compared to conventional CPU designs.

This matters strategically because the system was built under sustained U.S. semiconductor export controls that have denied Chinese institutions access to advanced GPU accelerators from Nvidia and AMD for AI-related purposes. LineShine demonstrates that a path to exascale performance exists without those components — though experts are careful to note that the benchmark where U.S. systems retain a decisive lead, HPL-MxP, is precisely the one most relevant to AI training workloads.

What Changes on the List

With LineShine’s arrival, the number of publicly verified exascale systems on the TOP500 rises to five. El Capitan, formerly ranked first, drops to second. Frontier falls to third. Aurora and JUPITER Booster round out the exascale tier in fourth and fifth place respectively. A notable new entrant is Italy’s HPC7, operated by Eni S.p.A., which debuts in sixth place with 571 petaflops. HPC7 uses the same HPE Cray EX255a platform and AMD Instinct MI300A architecture as El Capitan, making it effectively a scaled-down version of the American system — and giving Italy a competitive presence at the leadership edge of European HPC for the first time.

The June 2026 Top 10 as a whole reflects an unusual degree of architectural diversity: custom Chinese ARM-based processors, AMD APUs, Intel Xeon systems, Nvidia Grace Hopper nodes, Fujitsu A64FX chips, and cloud-based hybrid configurations all appear within the top tier. No single technology path dominates.

Scientific Applications

Since becoming operational, LineShine has supported research across a broad range of domains, including atmospheric and oceanic modeling, engineering simulation, materials science, drug discovery, neuroscience, scientific AI, and large-scale model inference. NSCS reports an average parallel scalability of 84.4% across mixed-precision computing workloads in large-scale parallel environments — a figure that speaks to the effectiveness of the system’s full-stack software co-design alongside its hardware.

The TOP500 list has been published twice yearly since 1993 and remains the authoritative public benchmark for high-performance computing leadership, even as some researchers note that the Linpack benchmark captures only one dimension of a system’s capabilities. For China, regardless of how the nuances are parsed, the message of LineShine’s debut is unambiguous: after nine years away from the summit, the country is back at the top.