On 12 August, SiEngine announced that its self-developed 7-nanometre automotive-grade AI accelerator, the Tiangong 100 (NNA100), has entered full mass production and begun batch supply to OEMs and Tier-1 partners. The chip is a standalone, plug-in accelerator for on-device agent AI, built to give local large models a complete domestic compute base and fix the shortfall in edge compute.
Its distinction is a standalone external AI accelerator that, SiEngine says, is the only one in China that needs no rebuild of the vehicle’s electrical and electronic architecture and stays compatible with all-brand cockpit main controllers. New or existing models can upgrade AI compute with minimal deployment, no changes to the original hardware or software, and far shorter delivery cycles.
On performance and safety, the Tiangong 100 delivers 96 TOPS INT8 with a dedicated 102 GB/s LPDDR5 memory bandwidth, isolating AI inference so it does not consume the vehicle’s base compute. It runs 3B to 7B models or several 4B models at once, and carries AEC-Q100 Grade 3 and ISO 26262 ASIL-B certification with a hardware security module protecting privacy.
SiEngine stresses that at an around 1,000-yuan hardware cost, with a highly open software-hardware ecosystem, the whole-vehicle deployment cost is at least halved versus high-compute dual-control AI box schemes. Hardware compatibility is broad, extending to SiEngine’s own Longying-1 and Xingchen-1 and mainstream domestic and foreign chips, covering entry to mid-to-high passenger platforms, and reaching industrial control and robotics.
Founder and chief executive Wang Kai said the goal is to help OEMs and Tier-1 partners clear the high barrier of large-model deployment and speed up adoption of 7B-and-above models across more vehicles and agents, pushing AI compute toward self-reliance.
Editor’s note: This is an adapted translation of the original Gasgoo report. It has been trimmed and restructured for readability for an international business audience.