🕷️
SetupAI
🐯Stalecollected in 11m

Li Auto's AI Chip-Model Co-Design Law

Li Auto's AI Chip-Model Co-Design Law
PostLinkedIn
🐯Read original on 虎嗅
#autonomous-driving#ai-chips#co-design#vla-modelsmach-100-chip

💡Math framework perfectly matches LLMs to AV chips, slashing costs & adaptation time

⚡ 30-Second TL;DR

What Changed

Introduces 'end-side LLM soft-hard co-design law' using loss scaling and Roofline modeling

Why It Matters

This enables AV firms to build scenario-optimal chips efficiently, accelerating edge AI deployment and challenging Nvidia dominance. It provides a replicable blueprint for the industry.

What To Do Next

Implement PLAS-like Roofline modeling to tune your edge LLM for Orin hardware constraints.

Who should care:Researchers & Academics

🧠 Deep Insight

Web-grounded analysis with 8 cited sources.

🔑 Enhanced Key Takeaways

  • •Li Auto mass-produced the first vehicles globally equipped with VLA models via the i8 launch in August 2025, expanding to the entire AD Max lineup by September 2025[1].
  • •Li Auto restructured its autonomous driving division into 11 sub-departments in 2025, subdividing algorithms into basic models, VLA models, and model engineering to support VLA as its core technology route[1].
  • •In March 2025, Li Auto unveiled the MindVLA architecture, a dual-system integrating end-to-end learning and Vision-Language Models, described by CEO Li Xiang as a 'robot large model'[2].

🔮 Future ImplicationsAI analysis grounded in cited sources

Li Auto's PLAS framework will enable deployment of optimized VLA models on next-gen chips like Mach 100 by late 2026
The framework targets the new L9 model with Mach 100 chip, aligning with Li Auto's 2026 growth plans and L9 updates featuring high computing power up to 2560 TOPS[6].
PLAS co-design will face regulatory hurdles from China's July 2027 Level 3 standards requiring minimal risk maneuvers and data recorders
New MIIT rules mandate advanced safety features for Level 3 systems, potentially impacting VLA deployment timelines in Li Auto vehicles[3].

⏳ Timeline

2025-03
Unveiled MindVLA architecture at Nvidia GTC
2025-08
Launched i8 as first global mass-produced VLA vehicle
2025-09
Expanded VLA to entire AD Max lineup and restructured AD division into 11 sub-departments
2026-01
Restructured R&D into foundation model, software, and hardware teams; Lang Xianpeng shifted to robotics
🐯Read original article on 虎嗅
📰

Weekly AI Recap

Read this week's curated digest of top AI events →

👉Related Updates

Same topic

Explore #autonomous-driving

Same product

More on mach-100-chip

Same source

Latest from 虎嗅

HarmonyOS ADS adds lane-level navigation and safety insurance

HarmonyOS ADS adds lane-level navigation and safety insurance

IT之家•Jul 1
Tesla Cybercab begins engineering tests in Austin

Tesla Cybercab begins engineering tests in Austin

36氪•Jul 1
Love and Deepspace faces backlash over new character design

Love and Deepspace faces backlash over new character design

虎嗅•Jul 1
AlphaGo's impact on Go and human players a decade later

AlphaGo's impact on Go and human players a decade later

虎嗅•Jul 1

AI-curated news aggregator. All content rights belong to original publishers.
Original source: 虎嗅 ↗