No 'universal solution' for inference chips: Scenario customization is the only path for automotive-grade standards
Community Discussion · Policy

No 'universal solution' for inference chips: Scenario customization is the only path for automotive-grade standards

Cockpit EnthusiastCockpit EnthusiastJul 182026/07/18 69 views

In 2025, China's frontier foundation models completed an iteration on average every two months, and model invocation costs already hold a clear advantage compared to similar foreign products. This is data revealed by Wang Dong, co-founder of Moore Threads, in his latest interview. What does this speed mean in the context of smart cockpits? It means we are experiencing a compute power iteration cycle faster than the smartphone era, yet the R&D cycle for automotive-grade chips starts at three years.

1 replies

?
Ctrl + Enter to reply
xiafeng
xiafengAug 1

This analysis is spot on. The modular combination approach has similar practices in the open-source community; for instance, using ONNX Runtime for heterogeneous scheduling can alleviate integration issues between different chips to some extent. There's a project on GitHub called "LlamaEdge" that is specifically optimized for automotive-grade edge devices, and its contribution guide is quite clear.