TL;DR: China's LineShine, a CPU supercomputer with 13,789,440 processor cores, posted 2.198 exaflops on the High Performance Linpack benchmark in the June 2026 TOP500 list, displacing the GPU-powered US system El Capitan (1.809 exaflops) as the world's fastest machine. LineShine is the first CPU supercomputer to break the 2 exaflop barrier without a single GPU accelerator, and the first Chinese system to lead the world rankings since Sunway TaihuLight in 2017.
What Just Happened at the June 2026 TOP500?
The 67th edition of the TOP500 list of the world's fastest supercomputers, announced on June 23, 2026 at the ISC High Performance conference (ISC 2026) in Hamburg, Germany, delivered a result that reshaped expectations across the high performance computing (HPC) industry. LineShine, a previously unlisted CPU supercomputer installed at the National Supercomputing Centre in Shenzhen (NSCS) and built by the Shenzhen Cloud Computing Center, posted 2.198 exaflops (2.198 quintillion floating point operations per second) on the High Performance Linpack (HPL) benchmark.
That score displaced El Capitan, the AMD-powered system at Lawrence Livermore National Laboratory (LLNL) in California, which recorded 1.809 exaflops on the June 2026 list. The margin is not narrow: the Chinese CPU supercomputer leads the second-place American machine by more than 20 percent, according to TOP500.org (June 23, 2026).
The result marks the first time a China-based system has led the TOP500 since Sunway TaihuLight last topped the list in November 2017, ending a nine-year run in which American and Japanese machines held the number one position. It also opens the CPU supercomputer era at the very top of the world rankings.
Who Built LineShine and Why Shenzhen?
The system was built by the Shenzhen Cloud Computing Center and installed at the National Supercomputing Centre in Shenzhen (NSCS), one of the network of national supercomputing centers China has operated since the early 2010s. NSCS has history on the TOP500: its Nebulae system reached the number two position worldwide in June 2010, at the start of China's first major push up the rankings.
The choice of Shenzhen is not accidental. The city is the hardware capital of China, home to Huawei, Tencent, and a dense supplier ecosystem for boards, cooling, and power electronics. Locating the machine there shortens every iteration loop between chip designers, system integrators, and the facility team, an advantage that matters when a new processor, a new interconnect, and a new platform all debut in the same system.
It also fits a broader pattern in Chinese compute policy: pairing each flagship machine with a regional industrial cluster. Wuxi anchored the Sunway program, Guangzhou hosted Tianhe-2, and Shenzhen now anchors the LingKun platform that LineShine introduces. Analysts quoted by Data Center Dynamics (June 2026) expect derivative LingKun systems at smaller scales to appear in provincial centers over the next two years.
What Is Inside the LineShine CPU Supercomputer?
LineShine is a CPU supercomputer in the strictest sense: it contains no GPU accelerators of any kind. No Nvidia H100. No AMD Instinct MI300X. Every floating point operation runs on domestically designed Chinese CPU silicon. The key specifications, as published by TOP500.org in June 2026:
- Processors: approximately 45,000 custom LX2 chips, each carrying 304 Armv9-based cores, for 13,789,440 cores in total
- HPL performance: 2.198 exaflops sustained, against a theoretical peak of 2.736 exaflops (roughly 80 percent efficiency)
- Interconnect: LingQi, a proprietary Chinese fabric built for the LingKun platform
- Operating system: Kylin, China's domestically developed Linux distribution
- Power draw: 42,220 kilowatts during the Linpack run, returning 52.07 gigaflops per watt
The 80 percent HPL efficiency figure deserves attention. Extracting four fifths of theoretical peak from a CPU supercomputer with almost 13.8 million cores requires an interconnect and a memory subsystem that keep every core fed. Many GPU-heavy systems run HPL at 65 to 75 percent of peak.
Energy efficiency is the trade-off. At 52.07 gigaflops per watt, LineShine sits well below the Green500 leader KAIROS at CALMIP (University of Toulouse-CNRS, France), which delivers 73.28 gigaflops per watt as of the June 2026 Green500 list. Running a CPU supercomputer at this scale means paying a real power bill: 42.2 megawatts is roughly the electricity demand of a mid-sized industrial park.
How Does the LineShine CPU Supercomputer Compare with El Capitan?
The table below compares the new number one with the two most recent American leaders, using figures from TOP500.org.
| System | Site | HPL performance | Architecture |
|---|---|---|---|
| LineShine | NSCS Shenzhen, China | 2.198 exaflops (June 2026) | CPU only (LX2, Armv9) |
| El Capitan | LLNL, USA | 1.809 exaflops (June 2026) | AMD EPYC CPUs + AMD Instinct MI300A |
| Frontier | ORNL, USA | 1.102 exaflops (debut, June 2022) | AMD EPYC CPUs + AMD Instinct MI250X |
Text takeaway: the CPU supercomputer LineShine outperforms El Capitan by 0.389 exaflops, a gap larger than the entire HPL score of most national flagship systems.
Why Does a CPU-Only Number One Matter?
For a decade, the prevailing assumption in HPC has been that peak performance requires GPU accelerators, and Nvidia's CUDA ecosystem became the default software stack for climate simulation, genomics, and large-scale AI training alike. A CPU supercomputer at number one breaks that orthodoxy, and it matters for three reasons.
Supply-chain independence
Using domestically designed processors rather than Nvidia or AMD accelerators removes exposure to US export controls, a central concern of China's national compute strategy since the October 2022 restrictions. A CPU supercomputer built on local silicon cannot be throttled by a foreign licensing decision. Readers following the policy angle can compare our analysis of AI model access risk in 2026, where similar dependency questions apply to model APIs rather than chips.
Workload flexibility
CPU-native HPC handles a broader class of workloads without code rewrites: legacy simulation codes, sparse matrix operations, financial risk models, and database-heavy analytics all run natively. Enterprises that standardize on a CPU supercomputer avoid the CUDA porting tax that GPU clusters impose on older codebases.
Architecture signal
Optimizing interconnects and memory bandwidth across 13.8 million cores shows that the performance gap with GPU-heavy designs can be closed by system engineering. Other builders of CPU supercomputer systems, including Fujitsu with the Arm-based Fugaku in Japan, have argued this for years. LineShine turns the argument into a benchmark result.
When Did China Last Lead the World Supercomputer Rankings?
China previously held the global number one position twice. Tianhe-2, built by the National University of Defense Technology in Guangzhou, led from 2013 to 2016 at 33.86 petaflops. Sunway TaihuLight, at the National Supercomputing Center in Wuxi, led from June 2016 through the November 2017 list at 93.0 petaflops. After that, American and Japanese systems dominated for nearly a decade until this CPU supercomputer restored China to the top.
| Period at number one | System | Country | HPL performance |
|---|---|---|---|
| 2013-2016 | Tianhe-2 | China | 33.86 petaflops |
| 2016-2017 | Sunway TaihuLight | China | 93.0 petaflops |
| 2018-2019 | Summit | USA | 148.6 petaflops |
| 2020-2021 | Fugaku | Japan | 442.0 petaflops |
| 2022-2024 | Frontier | USA | 1.102 exaflops (debut) |
| 2024-2025 | El Capitan | USA | 1.742 exaflops (debut, November 2024) |
| June 2026 | LineShine | China | 2.198 exaflops |
Text takeaway: top-end HPL performance has grown roughly 65-fold in thirteen years, from Tianhe-2's 33.86 petaflops in 2013 to the 2.198 exaflops the LineShine CPU supercomputer posts in June 2026.
What Are the Limits of the Linpack Benchmark?
HPL measures sustained double-precision (FP64) matrix arithmetic. It has been the yardstick of the TOP500 since 1993, and it rewards exactly the dense linear algebra that a CPU supercomputer with a strong interconnect can deliver. It does not measure the mixed-precision throughput that dominates modern AI training.
Industry analysts, including DigiTimes (June 24, 2026), caution that LineShine's Linpack crown does not imply Chinese leadership in AI compute, where Nvidia-class GPU clusters running FP8 and FP16 workloads still set the pace. A CPU supercomputer can lead the TOP500 while a GPU cluster next door trains larger language models faster.
There is also a transparency caveat. Chinese institutions stopped submitting many systems to the TOP500 after 2021, so the list understates China's installed base of CPU supercomputer capacity. LineShine's public debut may signal a policy change as much as an engineering one.
What Is the LX2 Processor Behind LineShine?
The LX2 is a custom Chinese server processor built on the Armv9 instruction set architecture, carrying 304 cores per chip. TOP500.org and ServeTheHome (June 2026) identify it as the compute engine of the LingKun platform, the system architecture underneath LineShine. At approximately 45,000 chips, the machine aggregates 13,789,440 cores, the largest core count ever recorded on the list.
A 304-core design signals an aggressive many-core philosophy. Rather than chasing the highest per-core clock speeds, the LX2 packs a very large number of moderate cores onto each package and relies on memory bandwidth and the LingQi interconnect to keep them busy. This is the same broad direction Fujitsu took with the A64FX processor in Fugaku (48 compute cores, Armv8.2 with SVE), scaled up by an order of magnitude in core count.
Several engineering details remain undisclosed as of early July 2026: the fabrication process node, the foundry, per-socket memory configuration, and sustained performance on non-Linpack workloads. The Next Platform (June 25, 2026) notes that the LX2's vector width and cache hierarchy will determine how well the CPU supercomputer generalizes beyond dense linear algebra, and those figures have not been published.
How Does This Change the US-China Compute Race?
Since October 2022, the United States has restricted exports of advanced accelerators and chipmaking equipment to China, with successive tightening rounds through 2023, 2024, and 2025 that covered Nvidia's A100, H100, and later the China-specific H20 line. The explicit goal was to slow Chinese progress in AI and high performance computing.
LineShine is China's most visible answer to date. By taking the number one HPL position with a CPU supercomputer that uses no restricted foreign accelerator, no US interconnect, and a domestic operating system, Beijing demonstrates that export controls raised the cost of Chinese HPC without capping it. The message is aimed as much at procurement officials in third countries as at Washington.
The response from the American HPC community will be watched at SC26 in November 2026, when the 68th TOP500 list is published. The US Department of Energy's post-El Capitan roadmap and Europe's JUPITER expansion at Forschungszentrum Julich in Germany both aim at the same 2-plus exaflop tier that the Chinese CPU supercomputer now occupies alone.
At the same time, the competitive picture is wider than one benchmark. US hyperscalers operate private GPU fleets whose aggregate training throughput dwarfs any single TOP500 entry, and none of those clusters submit HPL runs. The list measures what institutions choose to reveal, and LineShine is above all a deliberate reveal.
How Should Data Teams Benchmark Their Own Workloads?
The LineShine debate is, at bottom, a benchmarking debate, and enterprise teams can apply the same discipline at their own scale. Before choosing between CPU-first and GPU-first infrastructure, it pays to answer four questions with measurements rather than vendor slides.
- What precision does the workload actually need? Financial risk engines, actuarial models, and regulatory stress tests typically require FP64 accuracy, the territory where strong CPU nodes shine. Recommendation systems and language model inference tolerate FP16 or FP8, which favors accelerators.
- Is the code compute-bound or memory-bound? Sparse matrix operations, graph analytics, and large joins are usually limited by memory bandwidth, not raw flops. Adding GPUs to a memory-bound pipeline often buys little.
- What would porting cost? Legacy Fortran and C++ simulation codes, and most database engines, run on CPUs today. The engineering months needed to move them to CUDA or ROCm are a real cost line that belongs in the comparison.
- What is the all-in cost per result? Power, cooling, and depreciation matter at every scale. LineShine's 42,220 kilowatt draw makes the point vividly: the right metric is results per watt per dollar, not position on any single ranking.
Teams that run this exercise honestly often land on a mixed estate: CPU capacity for steady-state analytics and compliance workloads, accelerator capacity for model training bursts. That conclusion mirrors what the June 2026 TOP500 list shows at the extreme end, where a CPU-only machine and GPU-heavy machines coexist in the top three, each best at what it was designed for.
What Does the LineShine CPU Supercomputer Mean for Vietnam?
The LineShine result has practical implications for organizations in Vietnam and Southeast Asia evaluating compute infrastructure, a market we covered in our analysis of Vietnam's 7 billion USD AI data center boom.
First, CPU-based HPC remains highly competitive for data-intensive workloads: financial risk modeling, regulatory stress testing, large-scale data processing, and traditional scientific computing. The assumption that serious HPC requires GPUs should be re-examined workload by workload before capital is committed. In many of those cases a CPU supercomputer or a CPU-first cluster is the cheaper and simpler answer.
Second, architecture diversity is supply-chain resilience. The GPU shortage of 2022-2024 and ongoing semiconductor trade restrictions show that betting an entire compute roadmap on one accelerator vendor is a risk position, not a neutral default. A hybrid strategy that can fall back on CPU supercomputer capacity keeps projects moving when accelerator lead times stretch.
Third, open platforms keep options open. DataCore's HPC-OOD service, built on the Open OnDemand framework used by major US national laboratories and universities, is architecture-agnostic: it schedules CPU, GPU, and hybrid workloads through the same portal. Teams that want to experiment with CPU supercomputer style scaling can do so without changing their tooling, and DataCore's data products supply the structured Vietnamese market data those workloads consume.
Frequently Asked Questions About the LineShine CPU Supercomputer
What is the LineShine CPU supercomputer?
LineShine is a Chinese CPU supercomputer installed at the National Supercomputing Centre in Shenzhen. It posted 2.198 exaflops on the HPL benchmark in June 2026, making it the fastest supercomputer in the world on the 67th TOP500 list, according to TOP500.org.
Why is a CPU supercomputer taking first place significant?
It is the first machine to exceed 2 exaflops of sustained FP64 performance without GPU accelerators. Every previous exascale leader, including Frontier and El Capitan, relied on GPUs. LineShine proves a CPU supercomputer can reach the same tier using CPU cores alone.
Does LineShine mean China leads the United States in AI?
Not directly. HPL measures double-precision arithmetic, not the mixed-precision throughput used in AI training. Analysts at DigiTimes and elsewhere note that GPU clusters remain the workhorses of large-model training, so the CPU supercomputer crown is a statement about system engineering and supply-chain independence more than AI capability.
What processors does the LineShine CPU supercomputer use?
Approximately 45,000 custom LX2 processors, each with 304 Armv9-based cores, for a total of 13,789,440 cores. The chips are domestically designed, connected by the proprietary LingQi interconnect, and run China's Kylin operating system.
How much power does LineShine consume?
42,220 kilowatts (42.2 megawatts) during its Linpack run, which works out to 52.07 gigaflops per watt. That is efficient for a CPU supercomputer at this scale but below the 73.28 gigaflops per watt of the Green500 leader KAIROS, as of June 2026.
Sources
- TOP500.org: "LineShine Debuts at No. 1 as the TOP500 Enters a New Global Exascale Era", June 23, 2026.
- Tom's Hardware: "China tops the list of fastest supercomputers with a CPU-only behemoth, ending US champion El Capitan's reign", June 2026.
- Network World: "China's LineShine dethrones El Capitan as the world's fastest supercomputer", June 2026.
- The Next Platform: "A Deep Dive On China's LineShine All-CPU, Exaflops-Class Supercomputer", June 25, 2026.
- ServeTheHome: "Arm CPUs Take Number 1 in Latest Top500 List with Chinese LineShine", June 2026.
- Data Center Dynamics: "LineShine: All-CPU Chinese supercomputer named world's most powerful", June 2026.
- DigiTimes: "China's LineShine supercomputer tops TOP500, but AI lead remains unclear", June 24, 2026.
Để lại một bình luận
You must be logged in to post a comment.