Rockchip unveils Automotive AI BOX

Rockchip has put a more concrete shape around a problem many carmakers in China are now trying to solve: how to run large AI models inside the vehicle, not in a distant cloud. In a May 7 post from the Beijing auto show, the chip company said it had launched an Automotive AI BOX with ModelBest, the Beijing-based model company also known as 面壁智能. The pitch was direct: add a dedicated in-cabin compute module so automakers can run multimodal AI features locally, even when network conditions are weak or unavailable. That matters because today’s cockpit systems already juggle navigation, voice, cameras and infotainment on tightly budgeted processors. Rockchip said its AI BOX is designed as an independent compute center for the cabin, so large-model inference does not consume the main cockpit controller’s resources. The company listed weak-network use, privacy protection, interaction latency and bandwidth limits as the main pain points the product is meant to address. (rock-chips.com) ### What exactly did Rockchip launch? Rockchip said the box combines an RK3576M automotive-grade system-on-chip with an RK1828 AI coprocessor. The RK3576M handles data intake over Ethernet or camera feeds, plus preprocessing tasks such as 2D imaging, AI voice and CNN-based recognition, while also supporting display and agent framework deployment, according to the company’s description. (rock-chips.com) The RK1828 is the more eye-catching piece. Rockchip said the coprocessor includes high-bandwidth DRAM and supports on-device inference for 7-billion-parameter language and vision-language models, as well as 3B and 4B “Omni” models. The company said time-to-first-token can come in under 100 milliseconds and output can exceed 120 tokens per second. ### Why add a separate AI box instead of upgrading the main cockpit chip? (rock-chips.com) Rockchip said the module can connect to a cockpit domain controller over USB or GMAC as an external expansion unit for new models or retrofits. It can also be integrated more directly, with the AI coprocessor paired inside the cockpit controller over PCIe or USB. That gives automakers two routes: add AI capability to an existing architecture, or fold the hardware into a new design. (rock-chips.com) The company said that split-compute approach is intended to decouple the cockpit controller from the AI coprocessor through distributed computing. In practice, that means the AI BOX is being positioned less as a gadget and more as a dedicated accelerator for in-cabin model workloads. ### What kinds of in-car features is this meant to power? Rockchip listed a broad set of intended uses, including vehicle-manual Q&A, driving-habit assistance, sentry mode, welcome mode, multimodal interaction, intelligent assistants, occupant interaction with emotion recognition, multi-track audio separation and other local AI functions. (rock-chips.com) The company also said the design aims for a bill of materials in the roughly thousand-yuan range, while meeting automotive environmental requirements. ModelBest’s own materials fit that positioning. On its website, the company says it is focused on putting smaller, efficient multimodal models onto end devices and advertises a “pure edge” cockpit GUI agent for automotive use. ### Who is ModelBest, and what did the two companies say? Rockchip said it signed a deep technical cooperation agreement with ModelBest during the Beijing auto show. The two companies said they would combine Rockchip’s hardware platform with ModelBest’s edge multimodal models and automotive applications to speed engineering deployment, improve inference efficiency and stability, and lower the development threshold for automakers adopting edge AI. (rock-chips.com) (modelbest.cn) Lin Zhengyuan, Rockchip’s senior vice president, said edge AI is becoming a core driver of intelligent cockpits and is raising demands on chip architecture and software ecosystems. Lei Shengtao, ModelBest’s chief operating officer, said models and chips are the two basic elements of edge AI and that deeper coordination between them determines the upper limit of the user experience. (rock-chips.com) ### What comes next for this platform? Rockchip said the partnership is aimed at moving edge large models from adaptation and verification into productization and scale. The next visible checkpoint is whether automakers adopt the AI BOX as an add-on for existing cockpit systems or integrate its components into new vehicle programs shown after the Beijing auto show cycle. Rockchip’s May 7 announcement and ModelBest’s automotive product materials remain the clearest public references for that next step. (rock-chips.com)