Baidu Unveils 'Super-Node' Architecture, Touting 95% AI Chip Performance Gain

Baidu showcased its new Kunlunxin "Super-Node" at the Baidu World Conference 2025, detailing significant performance gains for its proprietary artificial intelligence hardware. The company claims the new system can boost single-card performance by 95% and increase single-instance inference performance by as much as eightfold.

According to company sources, the Kunlunxin Super-Node was fully enabled within Baidu's Baige 5.0 platform in April 2025 and has since been deployed at scale. The company demonstrated the architecture's capability by running a 1-trillion-parameter open-source model in a matter of minutes using a single cloud instance.

The performance increases are the result of integrating multiple Kunlunxin AI accelerator cards into a unified "Super-Node" architecture. This structure, when optimized under the DeepSeek V3/R1 PD split inference framework, yielded the 95% improvement in single-card efficiency. Baidu presented the public debut of the system as a showcase of its "core chip technology."