At the 2026 World AI Conference (WAIC), Rockchip exhibited under the theme “AIoT 2.0: Defining Next-Generation Intelligent Hardware,” showcasing its full stack of RK35xx host SoC plus RK1828 edge-AI co-processor dual-chip solutions covering mass-produced schemes for AINAS, AINVR, audio, vertical industries and in-vehicle AI boxes.
On-device model adaptation was the highlight. The RK3588/RK3576 + RK1828 dual-chip architecture has been deeply tuned at the hardware level for Gemma4, Qwen3/Qwen3.5 and Qwen3.5-Omni, running mainstream 2B-7B models smoothly on-device, with Gemma4’s first-token latency measured as low as 169.93 ms and strong multimodal interaction and agent performance.
Built around the RK182X co-processor series, Rockchip has established a complete “chip–model–device” chain: the host SoC handles general computing and multimedia, while the co-processor connects via PCIe to carry large-model inference, letting users upgrade AI capability without redesigning existing systems.
Feedback from the show floor suggests on-device LLMs are moving from demos to volume production, with clear demand from education, retail, automotive and robotics scenarios for locally runnable, cost-controlled AI — and the dual-chip architecture offers existing RK3588/RK3576 installations a smooth upgrade path.
Source: Rockchip official site