When the U.S. Department of Commerce announced restrictions on the maintenance of 7nm equipment, Chen Tian Shi was standing in the corridor of the Shanghai wafer fabrication plant. The monitoring screen displayed: the 92nd batch of Siyuan 590 chips passed the final inspection, with a yield rate fixed at 78.3%. The fist of this scientist from Jiangxi slowly relaxed—only 72 hours remained until the U.S. sanctions took effect. This 300-day life-and-death race is a microcosm of China’s breakthrough in AI computing power.
Chapter 1: The Genius Youth and the National Treasure
In 2009, in Beijing’s Zhongguancun, 26-year-old Chen Yunji made a crucial mark on the Longxin 3 architecture diagram while his brother Chen Tian Shi was debugging neural networks in the adjacent lab. This pair of prodigies from the University of Science and Technology, two years apart, one inheriting the spark of “Chinese chips,” the other immersed in the wilderness of artificial intelligence. Before AlphaGo was born, the brothers had foreseen the computing power revolution: “The CPU+GPU architecture will eventually be replaced by dedicated AI chips.”
In the spring of 2016, in the basement of the Beijing Computing Institute, the first deep learning processor, “Cambricon 1A,” lit up, but voltage fluctuations burned out three oscilloscopes. The team used an insulated box as an anti-static chamber and completed packaging tests through manual soldering. This chip, only the size of a fingernail, supported Huawei’s Mate 10 to achieve the world’s first AI photography on a mobile phone—China’s first time standing at the starting line in the AI chip field.
Chapter 2: A Counterattack Amid Supply Chain Storms
The Darkest Moment After Huawei’s Exit
In the deep autumn of 2019, when the termination of cooperation letter arrived at Huawei’s Shenzhen base, Cambricon’s revenue plummeted by 90%. The media predicted, “Without Huawei, it is undoubtedly doomed.” Chen Tian Shi held an emergency meeting overnight: “Cut off the IP licensing business, All in cloud chips!” Executives strongly opposed: “This is self-harm!” He remained silent for a moment and replied, “To survive, we must grow our own wings.”
The R&D team launched the “Xuan Si Plan” in a closed base in Hefei. The whiteboard in the corridor displayed a countdown: 300 days until the U.S. sanction window. Chip architect Li Wei’s diary recorded: “November 3, 7nm process simulation failed; December 17, memory bandwidth bottleneck breakthrough; March 9 of the following year, instruction set reconstruction plan passed…”
Siyuan 590: The Lightning of Sanction Night
On the eve of the tape-out in 2022, TSMC suddenly terminated cooperation. SMIC sent a 12-person task force to build a “virtual process line” at the Hefei base. Engineer Wang Haifeng innovated a mixed etching method to simulate a 7nm environment using a 14nm production line. The final delivered Siyuan 590 produced stunning results:
-
Computing Power Density: Training performance under unit power consumption reached 92% of A100.
-
Ecological Breakthrough: Self-developed MLU instruction set compatible with PyTorch native environment.
-
Energy Efficiency Miracle: Energy consumption for training trillion-parameter large models reduced by 37%.
When this chip lit up at the Nanjing Intelligent Computing Center, the U.S. Department of Commerce’s sanctions officially took effect—China won its first breath in the computing power battlefield.
Chapter 3: Three-Front Operations: The Computing Power Matrix of Cloud, Edge, and Terminal
Cloud: The Backbone of the National Computing Power Network
The Xuan Si 1001 cluster at the Kunshan Supercomputing Center is running the world’s largest Chinese AI “Wudao 3.0.” Its unique design has shocked the industry:
-
Dynamic Power Wall: Automatically adjusts computing power allocation based on grid load.
-
Cold Plate Liquid Cooling: PUE value reduced to 1.06, breaking the energy efficiency record of data centers.
By 2025, this type of intelligent cluster will be deployed in 23 cities, undertaking 40% of the national large model training tasks.
Edge: The Intelligent End of Industrial Pulse
At the CATL battery factory, the Siyuan 220 edge chip is performing an alternative legend:
-
10W Ultra-Low Power Consumption: Continuously analyzes ultrasonic data from 2000 battery cells.
-
Millisecond-Level Response: Defect detection speed increased by 50 times.
When German engineers visited, they marveled: “You have turned computing power chips into industrial sensors!”
Terminal: The AI Genes of Hundreds of Millions of Devices
The OPPO foldable phone hides the magic of the Cambricon 1M processor:
-
Privacy Guardian: Runs visual algorithms locally, photo editing with 0 data upload.
-
Battery Life Black Technology: AI call noise reduction power consumption is only 0.3 milliamps.
This “invisible chip” has been embedded in 370 million devices, becoming the largest source of AI terminal computing power in China.
Chapter 4: Ecological Breakthrough: The Expedition of the Wall Breakers
Neuware: The Tower of Babel for Developers
In 2024, Shanghai programmer Zhang Ming created miracles using Cambricon Neuware:
# Three lines of code cross cloud, edge, and terminal
model = load("resnet50")
model.compile(platform=["cloud", "edge"])
deploy(model, mobile_phone)This “write once, run everywhere” framework is attracting 280,000 developers to migrate. The secret lies in the elastic core of the MLUarch05 architecture:
-
Dynamic Instruction Set: Automatically adapts to TensorFlow/PyTorch operators.
-
Heterogeneous Scheduler: Seamlessly coordinates CPU/FPGA hybrid computing.
The Covert War in the Open Source Community
In 2025, one of the hottest projects on GitHub was the mysteriously open-sourced “Penglai Compiler” from Cambricon. Developers discovered its secrets:
-
Automatic Escape Technology: Converts CUDA code into MLU instructions.
-
Computing Power Recovery Technique: Utilizes idle units of NVIDIA graphics cards to accelerate Cambricon chips.
This “parasitic compatibility” strategy is quietly dismantling the CUDA ecological barrier.
Chapter 5: The Battlefield of Life and Death in 2025: Advancing Under the Watch of Giants
Capacity Strangulation War
When the price of NVIDIA H20 chips soared to 250,000 yuan at Shenzhen Port, Cambricon’s Hefei factory was undergoing a capacity revolution:
-
Photon Etching Breakthrough: 14nm production line achieves 7nm equivalent precision.
-
Chip Stacking Technology: Three Siyuan 370 packages achieve 590 performance.
These innovations increased monthly production capacity from 30,000 to 80,000 pieces, capturing 30% of domestic replacement demand.
The Jungle Law of the Billion-Dollar Market
The global AI chip market presents a brutal pyramid:
-
NVIDIA: 80% share builds a CUDA moat.
-
AMD + Intel: 11% share clings to traditional servers.
-
Cambricon: 1% share grows fangs in the cracks.
Its killer move lies in the data from the Nanjing Intelligent Computing Center: when the Siyuan 590 cluster runs the Baichuan large model, the training cost is only 63% of that of A100.
Chapter 6: The Future Map: The Stars and Seas of China’s Computing Power
The Normandy of Intelligent Driving
BYD’s “God Eye” intelligent driving system, equipped with the Siyuan 590X chip, set records under 70°C heat in Turpan:
-
Interference Resistance Algorithm: Maintains centimeter-level positioning in sandstorms.
-
Energy Efficiency Control: Total power consumption less than 1/5 of the vehicle’s air conditioning.
This chip has secured 38% of domestic intelligent driving orders, becoming Tesla’s nemesis in China.
Quantum-Classical Hybrid Computing Power
The most radical experiment in 2025 is unfolding in Hefei: Cambricon chips collaborate with “Origin Quantum” for drug simulation. When a task that traditionally requires 10,000 years is compressed to 3 days, Academician Yu Li of the Chinese Academy of Sciences wrote a note: “This is a wormhole to AGI.”
The Destiny of Chip Makers
Standing at the cargo terminal of the Hengqin Port, Chen Tian Shi gazes at the containers bound for Southeast Asia. Each standard container holds 50,000 Siyuan 370 chips—they will drive AI recommendations for e-commerce in Indonesia, CT image analysis in Thai hospitals, quality inspection robots in Vietnamese factories…
As the cargo ship sets sail with a horn in the twilight, the sound carries the salty sea breeze. This scene recalls that night in the basement in 2016, when burnt circuit boards released wisps of smoke. The young scientists could not have imagined that nine years later, these silicon-based lives would become the lifeblood of the intelligent world, flowing across 9.6 million square kilometers of land.
Just as the chip etching machine leaves physical marks on the wafer: All breakthroughs in core technology are the integration of suffering over time. When the Cambricon team finally saw the green micro-light of “Power On” after the 47th tape-out failure, China’s computing power long march had just crossed the Chishui River.