The World’s Most Powerful Supercomputer Has a New Owner

LingSheng, a supercomputer built in Shenzhen and led by a Sun Yat-sen University professor, has just taken the top spot on the global TOP500 supercomputing list, ending nearly a decade of Western dominance.

Even more striking: this No. 1 system uses no NVIDIA hardware and not a single GPU.

Among supercomputing systems above 2 EFlops, it stands alone, surprising observers abroad.

Specifically, LingSheng uses a fully CPU-based architecture designed for independent control, with its CPU chip, storage architecture and high-speed interconnect network all developed in-house.

Turing Award winner Jack Dongarra called LingSheng “a beacon of hope for a new AI4S architecture,” saying it will redraw the global computing power competition.

Zero GPUs, Full-Stack Independent Control

LingSheng does not use a single GPU, making it almost an outlier in today’s supercomputing field.

El Capitan, Frontier, Aurora: the exascale systems near the top of the current TOP500 all rely on GPUs for their main computing power.

GPUs excel at large-scale parallel computing and have been the core driver of supercomputing performance gains over the past decade.

LingSheng, however, chose a completely different path: an Online Acceleration all-CPU architecture, with chips, interconnects, storage, operating system and upper-layer software all developed in-house.

LingSheng’s core processor is the self-developed LX2 chip.

LX2 is not a general-purpose CPU in the traditional sense. It integrates an AI matrix acceleration unit and natively supports FP64, FP32, BF16, FP16 and INT8 precision formats, embedding matrix acceleration directly into the CPU.

Officials said LingSheng has surpassed traditional CPU+GPU heterogeneous architectures on key metrics such as mixed precision and sparse/dense matrix computing.

LX2’s memory architecture is another highlight of LingSheng.

LingSheng is the first supercomputing system to integrate domestically produced HBM. Each LX2 chip carries 32GB of domestic on-chip memory, delivering 4TB/s of memory bandwidth, while also supporting up to 256GB of external DDR5 memory to balance bandwidth and capacity.

Compared with traditional CPUs, LingSheng offers a 10-fold increase in memory bandwidth.

At the network layer, LingSheng uses the self-designed Lingqi high-speed interconnect network, supporting ultra-large-scale networking with 2 million ports and 100,000 nodes, with inter-node bandwidth reaching 1.6Tb/s.

On the software side, LingSheng runs a self-developed full-stack software platform that exposes the system’s low-level hardware capabilities to upper-layer applications in a programmable and optimizable way.

For cooling, LingSheng pioneers a 100% liquid-cooled cabinet design. The full system consumes 42.2MW and achieves an energy efficiency ratio of 52.07GFlops/W, ranking 50th on the Green500 energy-efficiency list.

LingSheng’s ability to top the TOP500 in double-precision floating-point performance using a pure CPU architecture, while also taking first place on the HPCG list, demonstrates the viability of the all-CPU path for traditional scientific computing workloads.

Its design, which embeds AI matrix acceleration directly into the CPU, also gives LingSheng a natural foundation for converged HPC and AI computing.

In large-scale parallel environments, LingSheng reaches an average scaling efficiency of 84.4%, with more than 10 million schedulable cores across the full system.

After Nine Years, China’s Supercomputing Returns to the Top

LingSheng’s win means China has returned to the top of the TOP500 after nine years. The last Chinese system to hold the No. 1 spot was Sunway TaihuLight in 2017.

And because LingSheng won with a fully domestic independent architecture, outside attention has focused on more than just its performance numbers.

LingSheng is the world’s first pure CPU-architecture supercomputing system to break 2 EFlops, challenging the industry convention that exascale computing power must depend on GPUs.

In supercomputing, GPUs have long held a dominant position.

Over the past decade, as demand for parallel computing surged, GPUs gradually replaced CPUs as the core source of computing power in supercomputing systems.

According to NVIDIA, more than 400 systems on the TOP500 list released at ISC2026 are powered by NVIDIA technology, accounting for 81% of the total.

The exascale systems near the top of today’s TOP500, including El Capitan, Frontier and Aurora, all use GPUs as their core computing engines.

GPU penetration in supercomputing has also made NVIDIA an almost unavoidable part of global computing power infrastructure.

The arrival of the AI era has further reinforced that structure.

Large model training’s demand for GPUs has bound NVIDIA’s supercomputing position tightly to the AI industry, creating an ecosystem barrier that is harder to shake.

LingSheng’s emergence is a rare exception within that structure.

At the awards ceremony, Turing Award winner Jack Dongarra said LingSheng showed the world a beacon of hope for supercomputing’s path toward a new system architecture for AI4Science.

LingSheng also topped the HPCG list, with performance of 22 PFLOPS, further validating its broad strength in traditional scientific computing workloads.

LingSheng’s chief architect, Lu Yutong, director of the National Supercomputing Center in Shenzhen and a professor at Sun Yat-sen University, has deep roots in the field.

From 2013 to 2015, Lu Yutong, then deputy chief architect of Tianhe-2, stood six times on the top podium of global supercomputing, witnessing Tianhe-2’s six-title streak.

Eleven years later, Lu returned to the stage to accept another award, with LingSheng carrying forward Tianhe-2’s history.

Beyond performance, LingSheng is already being used across multiple scientific fields.

AI for Science, or AI4S, is becoming one of the most closely watched frontiers in supercomputing.

From AlphaFold’s protein structure prediction to large-scale climate models, drug molecule screening and materials property prediction, the convergence of AI and traditional scientific computing is rapidly reshaping research paradigms across basic sciences.

This places new demands on supercomputing systems: they must not only run fast, but also support both high-precision scientific simulation and large-scale AI training, two very different types of workloads.

Existing GPU-accelerated architectures perform well in AI training, but at the architecture level they still face an inherent divide when deeply integrating the two computing modes.

LingSheng’s arrival helps fill that gap.

Since LingSheng was deployed, it has supported computing tasks in atmospheric and ocean science, engineering simulation, materials science, drug discovery, brain science, scientific AI and large model inference.

For large-scale application needs across science, engineering and industry, LingSheng provides a scientific intelligence application platform that integrates multiple disciplines, full workflows and multiple precision formats, and has already produced world-class application results.