VibeTimes
#기술

Without Cambricon, There Would Be No DeepSeek

모민철모민철 기자· 9/19/2026, 3:48:52 PM· Updated 9/19/2026, 4:53:07 PM

On September 18, Cambricon and Sugon (Haiguang Information) moved to strengthen their software ecosystems for AI chips. Even with an AI chip in hand, separate software is needed for AI models like DeepSeek to run on top of it. An AI accelerator—a semiconductor that processes computations quickly—does not immediately run AI models on its own. Just as an engine alone won't let you drive a car, software is required to convert and optimize models to fit the chip's architecture. Cambricon relies on NeuWare and MagicMind, while Sugon offers a platform that allows different types of chips to be used together, supporting AI models to run smoothly on their chips.

Where Nvidia solves this role with its CUDA ecosystem, Cambricon's pillars are NeuWare and MagicMind. NeuWare is a development platform spanning cloud, edge, and terminal processors, supporting AI model development, migration, and optimization. MagicMind analyzes models built in different frameworks and generates and optimizes backend code so they can perform inference on MLU products. Simplifying Cambricon's structure, it can be summarized as: 'DeepSeek model → NeuWare/MagicMind → MLU → AI service.' The MLU370 comes with a basic software platform and the MagicMind inference acceleration engine, providing an integrated architecture for both training and inference.

Cambricon's inference software has completed adaptation and optimization not only for DeepSeek but also for China's major large language models (LLMs), including Zhipu AI's GLM, Alibaba Cloud's Qwen, Moonshot AI's Kimi, and MiniMax. This demonstrates that how quickly various AI models can be adapted to one's own chips is key to the competition in AI accelerators.

Sugon's heterogeneous computing platform, based on CPUs and DCUs, supports major AI frameworks such as TensorFlow, PyTorch, and Caffe2. According to Sugon, it completed DCU adaptation for DeepSeek V3 and R1 in February 2025 and has since expanded support to newer models such as V3.2.

In the first half of 2026, Cambricon posted revenue of 5.996 billion yuan and net profit attributable to controlling shareholders of 2.311 billion yuan, up 108.13% and 122.61% year-on-year, respectively.

Related Articles