China’s LineShine becomes world’s fastest supercomputer without US chips

At a glance:

• LineShine achieved 2.198 exaflops on the LINPACK benchmark, surpassing the US Department of Energy’s El Capitan.
• The machine is the first Chinese system to top the TOP500 list since 2017, built entirely without Nvidia, AMD, or Intel chips.
• Its all-CPU design uses custom LX2 processors with Armv9 architecture, signaling a shift toward self-reliance amid US export controls.

A historic milestone in supercomputing

On Tuesday, the TOP500 organizers declared LineShine, a supercomputer housed at the National Supercomputing Center in Shenzhen, as the world’s fastest machine. The system reached 2.198 exaflops on the LINPACK benchmark, marking a milestone for China after a seven-year gap since its last number-one ranking. This achievement ends a prolonged dominance by US systems, including El Capitan, which had led since November 2024 and supports critical national security tasks like nuclear stockpile maintenance.

The victory is more than symbolic. LineShine’s performance represents a 20% leap over El Capitan, the previous leader, and underscores China’s growing capabilities in high-performance computing. For context, the TOP500 list has long been a barometer of technological prowess, with the US and Japan historically trading places at the top. LineShine’s rise signals a shift in the global balance of computational power, particularly as it bypasses the GPU-centric architectures that have defined recent supercomputing advancements.

A CPU-only design defying conventional wisdom

Most top-tier supercomputers today rely on graphics processing units (GPUs) from Nvidia and AMD to accelerate performance, especially for AI-driven workloads. LineShine, however, is an all-CPU machine, a design choice that reflects both necessity and innovation. Its custom LX2 processors, built on Armv9 architecture licensed from UK-based Arm Holdings, integrate GPU-style mathematical operations directly into the CPU cores. This approach eliminates the need for discrete GPUs, a strategic move given US export restrictions on advanced chips.

The machine’s scale is staggering: 14 million computing cores packed into 90 cabinets, consuming 42 megawatts of power. Despite its size, LineShine maintains efficiency at 52 gigaflops per watt, a notable feat for a system of this magnitude. It has already completed complex simulations, including Earth system modeling and human brain mapping, demonstrating its capability for scientific research. The use of KylinOS, a Chinese Linux variant, and the LingQi network further highlights the system’s homegrown stack, from software to interconnects.

Political messaging and self-reliance under export controls

LineShine’s submission to the TOP500 list carries political weight. China had ceased entering machines into the rankings in 2023 amid tightened US export controls on semiconductors. The decision to re-enter with LineShine was deliberate, according to its designers, who emphasized that the project was funded independently of government resources. This independence allowed them to bypass potential restrictions and showcase a fully indigenous system.

Experts see the move as a direct response to US policy. “They upped us by developing a system that is not reliant on GPUs,” said Jack Dongarra, a TOP500 organizer and leading HPC researcher. The timing aligns with broader trends: US restrictions on GPUs, crucial for AI, have inadvertently spurred China’s push for self-sufficiency in chips, software, and networking. Jimmy Goodrich of the University of California Institute on Global Conflict and Cooperation noted that the US regulatory framework has a loophole—CPUs face looser controls than GPUs, enabling systems like LineShine to flourish.

Limitations in the AI race

While LineShine dominates the traditional TOP500 benchmark, its performance on AI-specific tasks tells a different story. The list prioritizes double-precision floating-point operations, essential for scientific simulations but less relevant to AI training, which often uses lower-precision calculations. On the mixed-precision HPL-AI benchmark, LineShine ranked fourth, achieving 7.92 exaflops—a smaller margin over its main score compared to GPU-based systems. This gap reflects the lack of dedicated low-precision circuitry in its CPU design.

Moreover, the world’s largest AI clusters operated by xAI, Microsoft, Amazon, and Google are absent from the TOP500 rankings, as they are commercial rather than academic projects. “If the hyperscalers submitted their systems, this ‘world’s fastest’ would not crack the top five,” Goodrich observed. Thus, while LineShine’s crown is legitimate, it does not equate to supremacy in AI computing, where the US remains dominant.

Broader implications for global tech dynamics

LineShine’s success illustrates how export controls can accelerate rather than hinder technological development. Chinese firms, from automakers to cloud providers, are increasingly designing custom chips to reduce reliance on Nvidia. Even Nvidia itself is pivoting toward Arm-based CPUs, signaling a potential shift in the semiconductor landscape. The machine’s architecture—a blend of CPUs and integrated GPU-like operations—could inspire new hybrid designs in supercomputing.

However, challenges persist. China’s domestic GPU industry lags behind US counterparts by years, limiting its ability to compete in AI-specific workloads. Building a record-breaking system from homegrown parts is a triumph, but closing the gap in practical applications like AI training remains a hurdle. The broader narrative, though, is clear: policies aimed at constraining China’s progress are fostering a parallel tech ecosystem that operates beyond US oversight.

Conclusion: A crown with caveats

LineShine’s ascent marks a pivotal moment in the global supercomputing race, demonstrating China’s ability to innovate under pressure. While it falls short in AI-specific benchmarks, its existence challenges the effectiveness of US export controls and highlights the unintended consequences of technological isolationism. The machine’s legacy may lie not in its raw speed, but in its proof that self-reliance can yield groundbreaking results—even in the face of geopolitical barriers.