SiEngine ships an AI accelerator it says bolts on without touching the E/E architecture
The chipmaker put its 7-nanometer automotive AI accelerator into mass production and began volume supply, pitching it as a card that adds on-vehicle model compute to cars already sold.
SiEngine Technology said on August 12 that its automotive-grade 7-nanometer AI accelerator, named Tiangong 100 and designated NNA100, had entered full mass production and was shipping in volume to vehicle manufacturers and Tier 1 suppliers.
The company set the part against what it described as the prevailing approach, in which running larger models on a vehicle means replacing the main cockpit system-on-chip — costly, slow, and hard to apply to models already in the field. Tiangong 100 instead uses PCIe 4.0 and USB 3.2 in a plug-in card format, which SiEngine said required no change to the existing head unit’s hardware or software, so both new and in-service vehicles could have their AI compute upgraded.
The chip carries 96 TOPS of INT8 compute and 102 GB/s of dedicated LPDDR5 memory bandwidth. SiEngine said the inference resources were fully isolated and did not draw on the compute serving the center console, instrument cluster or vehicle control, and that the part supported models of three to seven billion parameters, or several four-billion-parameter models at once. It listed AEC-Q100 Grade 3 and ISO 26262 ASIL-B certification, a hardware security module, and a Cortex-A78AE functional-safety processor able to boot independently.
SiEngine described the part as the only automotive-grade AI accelerator in China requiring no rebuilding of the vehicle E/E architecture, a claim it makes for itself. It put hardware cost at the thousand-yuan level, and said the part cut solution cost by at least half against a high-compute dual-controller AI box.
Source: SiEngine Technology, August 12, 2026.
Source: SiEngine Technology