Analysis of Technical Route Differences Between Kunlunxin and Cambricon

Based on the searched data, both Kunlunxin and Cambricon occupy important positions in China’s AI chip sector, but they have significant differences in their technical routes:

Dimension	Kunlunxin	Cambricon
Architecture Type	Self-developed XPU-R architecture	ASIC dedicated architecture + DSA computing architecture
Instruction Set	Not fully self-developed instruction set	Fully self-developed Cambricon ISA instruction set
Technical Route	DSA architecture similar to Google TPU	Hybrid technical path similar to ARM and Google TPU [1]

• Origin: Internal AI accelerator project of Baidu, later developed independently
• Core advantage: Deeply optimized for Baidu’s internal scenarios such as search, recommendation, voice, and image [3]
• Application scenarios: Wide range of application cases in the Internet industry; the new product R200 is based on the new generation Kunlunxin self-developed architecture XPU-R [3]
• Software ecosystem: Deeply integrated with Baidu PaddlePaddle, and supports other mainstream frameworks [3]

• Positioning: Focus on the AI dedicated chip track, covering cloud, edge, and end [1]
• Product line:
  • Cloud: Si yuan 290, Si yuan 370 (up to 256 TOPS INT8 computing power)
  • Edge: MLU220 series
  • End: Intelligent processor IP [2]
• Technical features: Adopts self-developed DSA computing architecture, forming different technical routes from Tesla DOJO’s memory-computing integrated architecture and Birentech’s near-memory architecture [2]

• High computing power density
• Supports mixed-precision computing and sparse computing acceleration [3]
• Significant improvement in versatility and performance (new product R200) [3]

• Advantage: In specific application scenarios, computing power can be higher than that of GPU [2]
• Disadvantage: Versatility is relatively poor [2]
• The gap with NVIDIA is mainly reflected in versatility, but it has advantages in specific application scenarios [2]

• Kunlunxin
  : Relying on the Baidu ecosystem, it is more inclined to applications in the Internet industry and deeply integrated with the PaddlePaddle platform [3]
• Cambricon
  : Oriented towards a wider layout of intelligent computing centers, with important projects nationwide, including Nanjing Intelligent Computing Center [2]

Both belong to the

of domestic AI chips [3], but there are differences in technical implementation details:

• Cambricon is more “specialized”, with a fully self-developed instruction set and excellent performance in specific scenarios
• Kunlunxin, while ensuring dedicated acceleration capabilities, pays more attention to versatility and compatibility with mainstream ecosystems

From the perspective of industry trends, GPUs with strong versatility have more advantages during the rapid development period of large models, but when certain paradigms of large models are fixed, optimized dedicated chips may perform better [1]. This is also the basis for the long-term coexistence of the two technical routes.

---

[1] 36Kr - “寒武纪已成老登” (https://m.36kr.com/p/3600720464412681)
[2] 东方财富网 - “AI大模型与异构算力融合技术白皮书” (https://pdf.dfcfw.com/pdf/H3_AP202510141762072518_1.pdf)
[3] 电子发烧友 - “主流国产AI算力芯片全景图” (https://www.eet-china.com/mp/a440458.html)