China's first domestic GPU + neuromorphic chip hybrid inference system for large models released
国内首个国产 GPU + 类脑芯片大模型异构混合推理系统发布，较同类国产 GPU 算力集群性价比提升一倍以上
This system is a significant advancement in China's domestic AI hardware, offering a more cost-effective alternative to traditional GPU-only setups for large model inference.
A new system combining a domestic GPU and a neuromorphic chip for large model inference was released at the 2026 China Computing Power Conference. It was jointly developed by several Chinese companies and universities. The system, tested with Deepseek V4, offers more than double the cost-effectiveness compared to similar domestic GPU clusters and reduces operational costs by over 40%. It is designed for high-real-time scenarios like token factories and AI code generation, and provides a fully domesticized inference foundation for sensitive industries such as finance and security.
国内首个国产 GPU + 类脑芯片大模型异构混合推理系统发布，较同类国产 GPU 算力集群性价比提升一倍以上
IT之家 9 月 13 日消息，9 月 11 日至 13 日，2026 中国算力大会在河北廊坊举行。 由移动云公司联合中国电子科技南湖研究院、北京灵汐科技、上海天数智芯、清华大学、北京大学共同打造的 国内首个国产 GPU + 类脑芯片大模型异构混合推理系统正式发布 。 IT之家获悉，经 Deepseek V4 实测， 系统相较同类国产 GPU 算力集群性价比提升一倍以上 、业务运营成本降低 40% 以上。 据介绍，成果可广泛适配 Token 工厂、AI 代码生成、多智能体协同、智能制造等高实时性场景，同时面向金融、安防、通信等安全敏感行业，提供全栈国产化推理底座。 官方表示，这套系统的核心思路是“分而治之，协同增效”，将大模型做 PD / AF 拆分，不同模块部署到最擅长的芯片上。Attention 计算交给国产 GPU，发挥其通用计算优势。FFN（MOE 专家）延时敏感模块交给类脑芯片，利用其存算一体、片上大容量 SRAM 的架构优势。 通过自研模型编译器、高速互联协议和统一推理引擎，实现两类算力的任务拆解、协同调度、结果聚合 ，彻底突破传统单 GPU 推理的固有瓶颈。
