Cambricon
Last reviewed
Apr 26, 2026
Sources
24 citations
Review status
Source-backed
Revision
v6 · 3,994 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Apr 26, 2026
Sources
24 citations
Review status
Source-backed
Revision
v6 · 3,994 words
Add missing citations, update stale details, or suggest a clearer explanation.
Cambricon Technologies (Chinese: 寒武纪科技; stock ticker: 688256.SS) is a partially state-owned, publicly listed Chinese semiconductor company that designs artificial intelligence processors for cloud computing, edge computing, and terminal devices. Headquartered in Beijing, the company was founded in 2016 as a spin-off from the Chinese Academy of Sciences (CAS) and is often referred to as the "NVIDIA of China" due to its focus on building a comprehensive hardware and software ecosystem for AI workloads. Cambricon listed on the Shanghai Stock Exchange STAR Market in July 2020 and posted its first annual profit in 2025 after nine years of consecutive losses.
Cambricon's roots trace back to 2008, when the Chinese Academy of Sciences invested an initial 10 million yuan into a "brain-inspired computing" research project at the Institute of Computing Technology (ICT). The project aimed to create a processor architecture specialized for deep learning and neural network computation. Researchers at the ICT developed the DianNao family of neural network accelerators and published influential papers on domain-specific instruction set architectures for neural networks, including the seminal 2016 paper "Cambricon: An Instruction Set Architecture for Neural Networks" at the International Symposium on Computer Architecture (ISCA).
The Cambricon instruction set architecture (ISA) is a load-store architecture that integrates scalar, vector, matrix, logical, data transfer, and control instructions into 64-bit instruction words. It contains 64 general-purpose 32-bit registers and uses on-chip scratchpad memory instead of a vector register file for data storage. Benchmarks demonstrated that the Cambricon ISA achieves roughly 6.4 times shorter code length than GPU code, 9.9 times shorter than x86, and 13.4 times shorter than MIPS for neural network workloads.
Cambricon Technologies was officially incorporated on March 15, 2016, when the Cambricon research project was spun out of the Chinese Academy of Sciences as a commercial venture. The company's name references the Cambrian period, the geological era marked by the rapid diversification of life forms, reflecting the founders' vision that AI would drive a similar explosion of technological progress.
In its early years, Cambricon focused on licensing its neural processing unit (NPU) intellectual property. The company's breakthrough came in 2017 when Huawei adopted the Cambricon 1A processor IP for the Kirin 970 system-on-chip, which Huawei billed as the world's first smartphone processor with a dedicated NPU. The Kirin 980, released in 2018, featured a dual-NPU configuration also based on Cambricon technology. During 2017 and 2018, more than 98% of Cambricon's revenue came from intellectual property licensing to Huawei.
In August 2017, Cambricon raised $100 million in a Series A funding round with investors including Alibaba Group and Lenovo. A Series B round closed in June 2018, led by China Merchants China Direct Investments and SDIC Venture Capital Management, with participation from CITIC, Alibaba, Lenovo Capital, and CICC Capital. The Series B valued the company at $2.5 billion, making it one of China's most valuable AI startups at the time.
In May 2018, Cambricon launched the MLU100, its first chip designed for cloud computing, marking the company's pivot from IP licensing toward building its own standalone chip products.
The relationship with Huawei ended abruptly in 2019. When Huawei unveiled the Kirin 990, it replaced Cambricon's NPU design with its own proprietary Da Vinci architecture (also known as Ascend Core, or ASCE). This was a significant blow to Cambricon, as IP licensing revenue dropped sharply. The loss forced the company to accelerate its transition toward selling its own branded chips and accelerator cards for cloud and edge deployments.
Cambricon received what was described as the fastest-ever approval for a STAR Market IPO. On July 20, 2020, the company debuted on the Shanghai Stock Exchange STAR Market (ticker: 688256), pricing its shares at 64.39 yuan each and raising approximately 2.58 billion yuan ($369 million). The stock surged on its first day of trading, opening at 260 yuan and briefly reaching 295 yuan before settling around 212 yuan, roughly 230% above the offering price.
After years of consecutive losses, Cambricon began turning the corner in late 2024, reporting its first quarterly profit in Q4 2024. In 2025, the company posted its first full-year profit since founding: revenue reached 6.497 billion yuan (a 453% year-over-year increase), and net profit attributable to shareholders was 2.059 billion yuan. The gross margin for 2025 was 55.15%, and the company announced its first-ever dividend distribution of 632 million yuan.
Chen Tianshi (born 1985 in Nanchang, Jiangxi Province) is the CEO and co-founder of Cambricon Technologies. The son of an electrical engineer father and history teacher mother, he entered the University of Science and Technology of China (USTC) at age 16 through its elite Junior Class program for gifted youth. He earned his bachelor's degree in mathematics in 2005 and his doctorate in computer science in 2010, both before turning 25. Chen Tianshi serves as the company's chief executive and is the public face of the organization. As of late 2025, the Bloomberg Billionaires Index estimated his net worth at over $22 billion, making him one of the wealthiest individuals in China and one of the world's richest people under 40.
Chen Yunji (born 1983) is the older brother of Chen Tianshi and co-founder of Cambricon. He joined the Institute of Computing Technology at CAS at age 20 as its youngest member on the Loongson (Godson) processor development team. He obtained his PhD by 24 and served as a chief architect of the Godson-3 microprocessor. After co-founding Cambricon in 2016, Chen Yunji returned to academia shortly afterward. He currently serves as a full professor at the Institute of Computing Technology and heads the CAS Laboratory of Processors, where he continues research on neural network processor design.
Cambricon's earliest products were IP cores and standalone chips designed for mobile and edge devices.
| Product | Year | Process Node | Description | Key Specifications |
|---|---|---|---|---|
| Cambricon 1A | 2016 | N/A (IP core) | First commercialized deep learning processor IP; adopted by Huawei for the Kirin 970 | Optimized for mobile inference workloads |
| Cambricon 1H | 2017 | N/A (IP core) | Second-generation edge AI processor, available in 1H8 (low-power computer vision) and 1H16 (general-purpose) variants | Supports CNN, RNN, SOM, SVM, k-NN, k-Means, decision tree |
| Cambricon 1M | 2018 | TSMC 7nm | Third-generation edge AI chip; world's first terminal AI processor supporting on-device training | 2, 4, or 8 TOPS versions; 5 TOPS/W efficiency (INT8) |
The Machine Learning Unit (MLU) family represents Cambricon's line of data center AI accelerators. The chips are also referred to by their Chinese product name, Siyuan (思元). Each generation has advanced the architecture, process node, and memory configuration.
| Chip (MLU Name) | Also Known As | Year | Architecture | Process Node | Memory | INT8 Performance | TDP | Primary Use Case |
|---|---|---|---|---|---|---|---|---|
| MLU100 | Siyuan 100 | 2018 | MLUv01 | TSMC 16nm | 16 or 32 GB DDR4 | 128 TOPS (balanced) / 166.4 TOPS (high-perf) | 80W / 110W | Cloud inference |
| MLU270 | Siyuan 270 | 2019 | MLUv02 | N/A | DDR-based | 128 TOPS (INT8); 256 TOPS (INT4) | N/A | Cloud inference; video analytics |
| MLU290 | Siyuan 290 | 2021 | MLUv02+ | TSMC 7nm | 32 GB HBM2 (1,228 GB/s) | 512 TOPS | 350W | Cloud training and inference |
| MLU370 | Siyuan 370 | 2022 | Chiplet-based | 7nm | 48 GB LPDDR5 (614.4 GB/s) | 256 TOPS | N/A | Cloud training and inference |
| MLU590 | Siyuan 590 | 2024 | Advanced MLU | SMIC N+2 (7nm) | 80 GB HBM | ~15-20% above Siyuan 290 | N/A | Cloud inference (large-scale) |
| MLU690 | Siyuan 690 | 2026 (expected) | Next-gen | SMIC N+2 (7nm) | TBD | Targets NVIDIA H100-class performance | TBD | Cloud training and inference |
The MLU100, introduced in May 2018, was Cambricon's first cloud AI accelerator. Built on TSMC's 16nm process using the MLUv01 architecture, it operated at a base frequency of 1.0 GHz ("balanced mode") and a boost frequency of 1.3 GHz ("high-performance mode"). In balanced mode, the chip delivered 128 TOPS of INT8 performance and 64 TFLOPS of FP16 performance within an 80W power envelope. The MLU100 was available as a PCIe x16 accelerator card with either 16 GB or 32 GB of DDR4 memory.
Released in 2019, the MLU270 upgraded the architecture from MLUv01 to MLUv02 and introduced hardware-based data compression to improve effective cache capacity and memory bandwidth. It featured 16 processing cores and supported INT4 (256 TOPS), INT8 (128 TOPS), and INT16 (64 TOPS) precision modes, along with FP16 and FP32 mixed-precision support. The chip included a built-in video decoding unit capable of decoding 48 simultaneous 1080p30 video streams and 16 streams of video encoding, positioning it for video analytics workloads in data centers.
The MLU290, also marketed as the Siyuan 290, began mass production in January 2021 and was Cambricon's first chip fabricated on TSMC's 7nm process. It adopted the OAM (Open Accelerator Module) form factor and integrated 32 GB of HBM2 memory with 1,228 GB/s bandwidth. Performance reached 512 INT8 TOPS, 256 INT16 TOPS, and 64 CINT32 TOPS, with a TDP of 350W. The MLU290 introduced MLU-Link, Cambricon's proprietary chip-to-chip interconnect technology for multi-accelerator configurations. This was the company's first product positioned for large-scale AI training workloads alongside inference.
Launched in 2022, the MLU370 was Cambricon's first AI chip to incorporate chiplet technology, integrating 39 billion transistors on a 7nm process. It was also the first cloud AI chip to use LPDDR5 memory (48 GB at 614.4 GB/s bandwidth), which offered roughly 1.5 times the power efficiency of GDDR6. The MLU370 supported FP32, FP16, BF16, INT16, INT8, and INT4 precision formats, delivering up to 256 INT8 TOPS and 24 FP32 TFLOPS. The product family included the MLU370-S4, MLU370-X4, and MLU370-X8 accelerator cards, with the X8 variant targeting high-end training scenarios.
The Siyuan 590 (also called MLU590) represents a significant milestone as it is manufactured using SMIC's N+2 process, a domestic Chinese 7nm-class node that uses DUV lithography rather than EUV. This shift from TSMC to SMIC was driven in part by U.S. export controls that restricted Cambricon's access to advanced foreign fabrication. The chip entered volume production in Q3 2024 and features 80 GB of high-bandwidth memory. Performance is estimated at roughly 80% of NVIDIA's A100, with improvements of 15 to 20% over the Siyuan 290. The Siyuan 590 is optimized for inference workloads and is best suited for mid-size server clusters of 1,000 to 3,000 cards.
The Siyuan 690 is Cambricon's next-generation chip, designed to rival NVIDIA's H100 in performance. As of early 2026, the chip remains in the testing phase, with large-scale mass production potentially slipping to the second half of 2026. Like the Siyuan 590, it is expected to be fabricated on SMIC's N+2 (7nm) process. Cambricon has set a production target of up to 500,000 AI accelerators in 2026, a figure that would include both the Siyuan 590 and Siyuan 690.
Cambricon has invested heavily in building a software stack to support its hardware, recognizing that a competitive software ecosystem is essential for widespread adoption, similar to how NVIDIA's CUDA platform has been critical to NVIDIA's dominance.
NeuWare is Cambricon's unified software development platform, first introduced at the company's product conference in November 2017. It provides a full-stack environment for deploying machine learning models on MLU hardware, integrating compilers, runtime frameworks, model optimization tools, and SDKs. NeuWare supports deployment across cloud, edge, and terminal devices. The platform has been positioned as a domestic alternative to NVIDIA's CUDA ecosystem, and Cambricon has described it as fully mature, underscoring China's push for software self-reliance in AI infrastructure.
BANG C is Cambricon's proprietary programming language for the MLU platform, analogous to CUDA C for NVIDIA GPUs or HIP for AMD GPUs. It supports both general-purpose computation and high-performance machine learning operations. BANG C exposes the MLU's specialized memory hierarchy, including separate Neuron RAM (NRAM) and Weight RAM (WRAM) spaces, allowing developers to write optimized kernels that take advantage of the hardware's architecture. The language is supported by the CNCC (Cambricon Neural Network C Compiler) and integrates with the broader CNToolkit development environment.
CNToolkit is Cambricon's comprehensive development toolkit, comparable to NVIDIA's CUDA Toolkit. It includes the following components:
| Component | Description |
|---|---|
| CNCC | Cambricon Neural Network C Compiler for BANG C programs |
| CNAS | Cambricon assembler |
| LLVM-MM | LLVM-based compilation backend for MLU targets |
| CNRT | Cambricon Runtime for executing compiled programs on MLU hardware |
| CNGDB | Debugger for MLU programs |
| CNPerf | Performance profiling tool |
| CNStudio | Integrated development environment |
| CNCodec | Media codec library for video encoding/decoding |
| CNPAPI | Performance application programming interface |
| CNDev / CNDrv | Low-level device driver and management interfaces |
MagicMind is Cambricon's inference acceleration engine, designed to convert trained models from popular AI frameworks (such as PyTorch, TensorFlow, and ONNX) into optimized formats for execution on MLU hardware. It performs graph-level and operator-level optimizations, including operator fusion, memory planning, and precision calibration. MagicMind is available as both a standalone tool and through GitHub as an open-source project (magicmind_cloud).
Cambricon operated at a loss for every year from its founding in 2016 through 2024, accumulating roughly 5 billion yuan in total losses. The company's financial trajectory shifted dramatically in 2025 as demand for domestic AI chips surged in China.
| Year | Revenue (RMB) | Net Income (RMB) | Notes |
|---|---|---|---|
| 2017 | ~111 million | Loss | Primarily IP licensing revenue from Huawei |
| 2018 | ~444 million | Loss | IP licensing revenue from Huawei (Kirin 970/980) |
| 2019 | ~444 million | Loss | Huawei partnership ended; pivot to cloud chips |
| 2020 | ~441 million | Loss | IPO on STAR Market |
| 2021 | ~721 million | Loss | MLU290 mass production began |
| 2022 | ~729 million | Loss | Added to U.S. Entity List |
| 2023 | ~709 million | -848 million | Seventh consecutive annual loss |
| 2024 | ~1.17 billion | -452 million | First quarterly profit in Q4; losses narrowed 47% |
| 2025 | 6.497 billion | 2.059 billion | First annual profit; 453% revenue growth YoY |
The 2025 results were driven almost entirely by the cloud product line, which generated 6.477 billion yuan in revenue (a 455% increase), while edge product revenue declined 48%. R&D expenditure in 2025 was approximately 1.2 billion yuan, down to 18% of revenue from 91% in prior years, a shift caused by the explosive revenue growth rather than reduced research investment.
Cambricon's stock (688256.SS) has been one of the most volatile on the Shanghai exchange. After its strong IPO debut in July 2020, the stock declined significantly over the following years as the company continued to post losses. By September 2023, shares were trading at roughly half of the IPO price.
The stock reversed course in 2024, surging 383% for the year and becoming China's best-performing stock of 2024, outperforming even NVIDIA and TSMC. By year-end 2024, Cambricon's market capitalization had reached approximately $37 billion. The rally continued into 2025, and by October 2025, the stock had risen further as the company reported its first profit. As of early 2026, shares traded between 520 and 1,596 yuan over the trailing 52 weeks, with a market capitalization fluctuating around 450 to 600 billion yuan ($60 to $80 billion).
The stock's rise has been driven by several factors: U.S. export controls limiting Chinese companies' access to NVIDIA's high-end GPUs, growing Chinese government mandates for domestic AI chips in data centers, and Cambricon's own operational improvements.
In December 2022, the U.S. Department of Commerce's Bureau of Industry and Security (BIS) added multiple Cambricon entities to the Entity List, citing the company's role in acquiring U.S.-origin items in support of China's military modernization. Cambricon received a "Footnote 4" designation, a novel Entity List category created in October 2022 that extends U.S. export controls to wholly foreign-made items if they are developed or produced with U.S. technology and shipped to a Footnote 4 entity.
Paradoxically, the U.S. restrictions have benefited Cambricon in some respects. By cutting off Chinese companies' access to NVIDIA's most advanced GPUs (including the A100 and H100), the export controls created a massive domestic market opportunity for Chinese chipmakers. Chinese cloud providers and AI companies have increasingly turned to domestic alternatives like Cambricon and Huawei Ascend processors. The restrictions also accelerated Cambricon's transition from TSMC fabrication to SMIC's domestic foundry processes.
Cambricon competes in a rapidly evolving Chinese AI chip market. As of 2024, the approximate market share distribution for AI chips in China was:
| Company | Market Share (2024) | Chip Family | Notes |
|---|---|---|---|
| NVIDIA | ~66% | A100, H100 (restricted) | Dominant but facing tightening export controls |
| Huawei (HiSilicon) | ~23% | Ascend 910, 910B | Largest domestic competitor; integrated ecosystem |
| AMD | ~5% | MI series | Limited presence in China |
| Cambricon | ~1% | Siyuan 590 / 690 | Growing rapidly; positioned as cloud inference specialist |
| Others | ~5% | Moore Threads, MetaX, Biren, etc. | Various stages of development |
Huawei remains Cambricon's most significant domestic rival. Huawei's Ascend product line benefits from a much larger market share (roughly 23 times that of Cambricon in 2024), a vertically integrated ecosystem spanning hardware to cloud services, and established relationships with major Chinese AI companies including DeepSeek and ByteDance. Huawei has also published an aggressive multi-year roadmap, with the Ascend 950 expected in 2026 targeting 1 PFLOPS at FP8 with 128 to 144 GB of memory.
Cambricon differentiates itself through its focused approach on specialized AI accelerator silicon and its open software tools. While Huawei builds a closed vertical stack, Cambricon positions itself more like NVIDIA, providing chips and a software platform that can be integrated into various server and cloud configurations. The Siyuan 590 is optimized specifically for inference workloads in mid-scale deployments, a niche where Cambricon claims competitive performance-per-watt advantages.
Cambricon's chips remain behind NVIDIA's latest offerings in absolute performance. The Siyuan 590 achieves roughly 80% of the performance of NVIDIA's A100 (a chip first released in 2020), while NVIDIA has since moved to the H100, H200, and Blackwell architectures. The Siyuan 690, if it reaches production targets, would approach H100-class performance. However, within China's restricted market, where access to NVIDIA's most advanced chips is limited, Cambricon's products fill a genuine need.
One of Cambricon's most pressing challenges is chip fabrication. The company depends on SMIC's N+2 (7nm-class) process for its Siyuan 590 and 690 chips. This process uses older DUV (deep ultraviolet) lithography rather than the EUV (extreme ultraviolet) lithography that leading foundries like TSMC and Samsung use for their most advanced nodes. The result is lower yields: reports from Bloomberg indicate yield rates of approximately 20% for Cambricon's largest dies, meaning four out of five chips fail to meet specifications.
Cambricon's ambitious 2026 production target of 500,000 chips would require SMIC to allocate significant capacity. Limited HBM (high-bandwidth memory) supply from domestic sources adds another constraint, as international HBM suppliers face their own export control restrictions when selling to Chinese chipmakers.
Cambricon Technologies is headquartered at the Zhizhen Building in the Haidian District of Beijing. The company maintains additional offices in Shanghai, Shenzhen, Hefei, Xi'an, Nanjing, Suzhou, and Zhuhai, along with a subsidiary in Hong Kong. As of the most recent available data, the company employs approximately 980 people. The Chinese Academy of Sciences remains a significant shareholder through its investment entities.
The company's full legal name is Cambricon Technologies Corporation Limited (中科寒武纪科技股份有限公司). The "中科" (Zhongke) prefix references its origin at the Chinese Academy of Sciences.
| Year | Event |
|---|---|
| 2008 | CAS initiates brain-inspired computing research project with 10 million yuan investment |
| 2016 | Cambricon Technologies incorporated on March 15 as CAS spin-off; launches Cambricon 1A IP |
| 2017 | Huawei adopts Cambricon 1A in Kirin 970; Series A raises $100 million; NeuWare software platform launched |
| 2018 | Cambricon 1M edge chip and MLU100 cloud chip released; Series B values company at $2.5 billion |
| 2019 | MLU270 released; Huawei ends partnership, develops own Da Vinci architecture |
| 2020 | IPO on Shanghai STAR Market raises $369 million; shares surge 230% on debut |
| 2021 | MLU290 (Siyuan 290) begins mass production on TSMC 7nm |
| 2022 | MLU370 (Siyuan 370) launched with chiplet design; U.S. adds Cambricon to Entity List |
| 2024 | Siyuan 590 enters volume production on SMIC 7nm; stock rises 383%; first quarterly profit in Q4 |
| 2025 | First annual profit: 2.059 billion yuan on 6.497 billion yuan revenue; first dividend announced |
| 2026 | Siyuan 690 in testing; production target of 500,000 chips for the year |