TL;DR
On September 12, 2026, at the 2026 China Computing Power Conference, China Mobile Cloud, CETC Nanhu Research Institute, Lynxi Technology, Shanghai Tianshu Zhixin, Tsinghua University and Peking University unveiled China's first domestic GPU + neuromorphic-chip heterogeneous hybrid inference system for LLMs. It routes Transformer Attention to domestic GPUs and FFN modules to neuromorphic chips, and was validated on DeepSeek V4 Flash with 2x+ inference output and energy efficiency and 40%+ lower operating cost.
Background: dual bottlenecks of LLM inference
As AI shifts from training to large-scale inference services, inference compute has become new digital infrastructure for token factories, AI agents and text-to-video services. Concurrent LLM serving imposes extreme demands on throughput and energy cost. Traditional single-GPU inference faces two bottlenecks: frequent data movement creates the "memory wall" and "power wall," while limited access to advanced domestic chip process nodes constrains gains from hardware scaling alone.
Approach: division of labor between GPU and neuromorphic chip
The project starts from system-architecture innovation, deeply integrating domestic general-purpose GPUs with neuromorphic chips for the first time, implementing a PD/AF split-inference mode for Transformer models:
- Attention computation: handled by domestic GPUs.
- FFN (feed-forward network) modules: handled by neuromorphic chips, leveraging compute-in-memory and large on-chip SRAM to break the inherent bottlenecks of GPU-only inference.
The team also built a self-developed model compiler, high-speed interconnect protocol and unified inference engine for task decomposition, coordinated scheduling and result aggregation across the two compute types.
Validation: tuned on DeepSeek V4 Flash
The system was fully tuned and validated on the DeepSeek V4 Flash LLM. Compared with a comparable domestic GPU cluster:
- Inference output and energy efficiency: both more than 2x;
- Operating cost: reduced by over 40%.
This shows that domestic chips on mature process nodes, combined with system-architecture innovation, can match next-generation international inference architectures, providing a replicable paradigm for domestic LLM inference.
Ecosystem collaboration and rollout
This result is a typical industry-academia-research collaboration, combining an operator, chipmakers, research institutes and universities across the full chain from chip adaptation and inference-engine development to scenario validation. The team will keep iterating toward token factories, AI coding, text-to-video, multi-agent collaboration, smart manufacturing and financial/telecom security, and plans to gradually open the core heterogeneous inference framework to domestic GPU and neuromorphic-chip vendors and computing operators.
Data sources & references
- CNR News (September 13, 2026): China Mobile's new heterogeneous hybrid LLM inference system debuts at 2026 Computing Power Conference
- People's Daily (September 12, 2026): 2026 China Computing Power Conference opens in Langfang, Hebei