China Takes Supercomputer Crown From U.S. for First Time Since 2017

 

 

“A supercomputer in Shenzhen was declared the world’s fastest. It uses only standard microprocessors and not the special-purpose chips called graphics processing units.

 

China took back a coveted computing crown from the United States on Tuesday, ratcheting up a fierce technological competition that has implications for science, national security and geopolitics.

 

LineShine, a massive computing system in Shenzhen, China, was declared the world’s fastest by a group of researchers using a set of standard tests for supercomputers. Besides raw speed, the system stood out because it uses only standard microprocessors and not the special-purpose chips called graphics processing units, which most high-end supercomputers rely on for heavy number crunching.

 

That underlying design could point to a better way to blend artificial intelligence with traditional scientific tasks, said Jack Dongarra, an organizer of the so-called Top500 list of the world’s most powerful supercomputers.

 

Dr. Dongarra, a professor of computer science and electrical engineering at the University of Tennessee, recently inspected the new machine, at the Shenzhen Cloud Computing Center. LineShine’s test results were more than 20 percent faster than those of El Capitan, a system at Lawrence Livermore National Laboratory in California that has topped a twice-yearly ranking of supercomputer performance since November 2024. China had not placed a machine at the top of the list since 2017.

 

“It’s an impressive system,” Dr. Dongarra said of LineShine. “They upped us by developing a system that is not reliant on GPUs.”

 

The new supercomputer adds to the race between China and the United States for technological supremacy. U.S. tech giants like OpenAI, Anthropic and Google have developed leading A.I. models, while another American company, Nvidia, has become the world’s dominant supplier of A.I. chips.

 

China has tried to innovate in different ways, with the Chinese start-up DeepSeek releasing a cutting-edge A.I. model last year using just a tiny fraction of specialized A.I. chips.

 

To prevent China from catching up, President Trump has imposed tariffs and at times placed limits on A.I. chip exports. But China’s use of standard microprocessors, which are known as CPUs, rather than GPUs to create an ultrafast supercomputer suggests a potential way to get around those roadblocks.

 

“The U.S. government should have stronger controls on the export and manufacturing of CPUs for the China market,” said Jimmy Goodrich, a senior fellow at the University of California Institute on Global Conflict and Cooperation. “It is a loophole in the current regulations.”

 

Supercomputers, a term for the largest machines dedicated to science, have been used since the 1960s for tasks like creating climate models, cracking codes and designing nuclear weapons. They typically use high-precision mathematics, expressing numbers with 64 bits of data.

 

Commercial A.I. systems from companies like Google and OpenAI, by contrast, can be even faster. They can use approximations for tasks such as identifying images or selecting the next word in a sentence, relying on what are known as four-bit and eight-bit numbers that allow the systems to make many simpler calculations at once.

 

“It is notable and impressive what China has done here, but they can’t hold a candle to these massive A.I. supercomputers that have been built by American A.I. labs” and others, Mr. Goodrich said.

 

U.S. national labs, which are the main buyers of some of the largest supercomputers, are eager to use A.I. to accelerate aspects of their scientific work. So they are adopting more of these less precise calculations, along with the 64-bit variety.

 

Though U.S. companies have historically dominated the ranks of the very largest supercomputers, foreign systems have sometimes vaulted to the top. A system in Japan, for example, ranked No. 1 on the list from 2020 to 2022.

 

“There’s a lot of talk that America is the only country capable of these systems,” said Addison Snell, an analyst at Intersect360 Research, a firm tracking the sector. “Then you find that other companies have capabilities, too.”

 

Powerful systems from China and Japan have regularly spurred the Department of Energy and other U.S. agencies to push for more funding for supercomputers. In November, the Trump administration started the Genesis Mission, which aims to exploit supercomputers at U.S. national labs, along with private companies, to supercharge A.I. and scientific research.

 

GPUs, primarily developed by Nvidia and Advanced Micro Devices, have been a critical weapon in the recent supercomputer race. These chips excel in doing many chores simultaneously, including so-called vector calculations used in science and matrix multiplication [1] at the heart of many A.I. tasks.

 

When U.S. officials limited China’s access to GPUs and other powerful chips, as well as restricting exports of some machines for manufacturing the most advanced semiconductors, that caused it “to invest in developing architectures and technology to effectively have supercomputers that are at the same level as the U.S.’s highest-performing systems,” Dr. Dongarra said.

 

China’s LineShine system does not separate the traditional jobs of microprocessors and GPUs, as most high-end systems do. Instead, it builds in GPU-style tasks with specialized circuitry that accelerates matrix and vector calculations. That ability is embedded in chips that have a total of nearly 14 million computing cores, or tiny electronic brains, installed in 90 hardware cabinets.

 

These chips are an original design based on a set of instructions licensed by Arm Holdings, a British company that is controlled by the Japanese conglomerate SoftBank. Arm’s technology is best known for powering smartphones but has lately been adapted by Nvidia, Amazon, Qualcomm and others for use in data centers.

 

Arm has long operated in China. “Arm operates globally, including in China, in compliance with applicable export control laws and regulations,” a company spokeswoman said.

 

LineShine’s designers, who are supercomputer veterans in China, have not disclosed details about which company manufactured the chips or the level of chip production technology used, Dr. Dongarra said.

 

He and other experts have long thought that China had systems capable of a No. 1 ranking, but laboratories there had not recently submitted test results.

 

“It doesn’t surprise me that there is a Chinese machine capable of being No. 1,” Mr. Snell said. “The surprise is that they wanted the acknowledgment.”

 

Dr. Dongarra, who wrote a detailed report on the new system, was told while visiting China that the system had been made without government funding, so the designers felt it was permissible to submit tests for the Top500 ranking, he said.

 

The Shenzhen scientists have also sought recognition for the new machine through 14 submissions for the Gordon Bell Prize, which promotes solving sophisticated problems in science, Dr. Dongarra said. Three systems are finalists for that award, and three for a related prize in climate science.

 

LineShine has been used for projects like a sophisticated simulation of Earth, including atmosphere, ocean, land and ice components, as well as a complex simulation of the human brain, according to Dr. Dongarra’s report.” [2]

 

1. Matrix multiplication is the foundational mathematical engine driving modern artificial intelligence. In deep learning, virtually all data—whether images, text, or audio—is converted into numerical grids called matrices. By multiplying these matrices against internal "weights," AI models learn patterns, make predictions, and generate content.

Why Matrix Multiplication Powers AI

At the microscopic level, a neural network consists of thousands or billions of simulated neurons. Each neuron receives inputs, multiplies them by specific learned weights, and passes them to the next layer.

Instead of processing these neurons one by one, developers use matrix multiplication to calculate the outputs for entire layers at once. This process, known as vectorization, condenses millions of individual calculations into a single, massive mathematical operation.

Ubiquitous AI Tasks

Different AI architectures rely on matrix multiplication in various ways:

•           Large Language Models (LLMs): Models like ChatGPT or Llama rely on transformer blocks, which use massive matrix multiplications to power "attention mechanisms". In these mechanisms, the AI mathematically evaluates the context of words—converting them to tokens, embeddings, and predicting subsequent sequences. Embeddings in machine learning are mathematical representations of real-world data—such as text, images, or audio—translated into lists of numbers (vectors). By placing similar objects close together in a continuous geometric space, embeddings allow AI models to understand semantic meaning, inherent properties, and relationships. Over 83% of an LLM's computational runtime is spent performing these operations.

•           Computer Vision: In visual tasks, an image is essentially a matrix of pixel values. Matrix multiplications—acting as "convolutions" or linear transforms—allow the AI to filter images, detect edges, and recognize objects.

•           Backpropagation: During the model training phase, matrix operations calculate gradients across the entire network backwards, allowing the AI to adjust its internal parameters and "learn".

The Marriage of AI and Hardware

Because matrix multiplication involves calculating the values for each cell in an output grid independently, it is highly parallelizable.

•           GPUs: Graphics Processing Units contain thousands of small cores designed specifically to perform parallel matrix math, executing billions of multiplications per second.

•           Optimized Hardware: Specialized hardware modules like NVIDIA's Tensor Cores are engineered specifically to accelerate these matrix multiply-accumulate operations.

The Search for Faster Algorithms

Because matrix multiplication is so computationally intensive, even minor improvements to the underlying algorithms yield massive performance and energy savings across data centers. The industry continues to push the envelope in this space.

 

2. China Takes Supercomputer Crown From U.S. for First Time Since 2017. Clark, Don.  New York Times (Online) New York Times Company. Jun 23, 2026.