At the "Computing Power First Release" session of the 2026 China Computing Power Conference on September 13, China Mobile Cloud unveiled the country's first large-model heterogeneous hybrid inference system combining domestic GPUs and neuromorphic chips.
Developed with CETC Nanhu Research Institute, Beijing Lingxi Tech, Shanghai Tianshu Zhixin, Tsinghua and Peking universities, the system splits Transformer computation: Attention to domestic GPU, FFN modules to neuromorphic chips, leveraging in-memory computing and on-chip SRAM to break traditional GPU bottlenecks.
On the DeepSeek V4 Flash model, versus comparable domestic GPU clusters, inference output and energy efficiency more than doubled while operating costs dropped over 40%. It has formed 15 authorized invention patents and 6 software copyrights and entered small-batch trial production.
The breakthrough offers an autonomous, controllable new compute path for Token factories, multi-agent collaboration and security-sensitive sectors such as finance and communications, benchmarking Nvidia's next-generation heterogeneous inference architecture.
Source: Tencent News / Cailian Press
