Li Auto is exploring an early-stage custom cloud-inference chip using the dataflow architecture of its autonomous-driving chip.

Li Auto is exploring the development of a proprietary cloud-inference chip, according to multiple sources cited by LatePost (晚点). The project remains at an early stage and is expected to use the same dataflow architecture as the company’s autonomous-driving chip.

The proposed processor is a cloud inference accelerator, distinct from the in-vehicle autonomous-driving chip whose architecture it would reuse. Cloud inference processors can handle selected workloads otherwise run on GPUs, including autonomous-driving model data processing, testing and simulation, as well as requests to large language models.

The report distinguishes inference from model training that involves parameter updates. Those parameter-update workloads would still require training processors, rather than being assigned to the proposed inference chip.