Aixin's M97 Chip Drives AI Growth

Post LinkedIn

🔥Read original on 36氪

#ai-chip #adas #edge-computingm97aixin-yuanzhi m97 vla vlm

💡700T Chinese ADAS chip leads bandwidth for VLA, launches Q3 2024.

⚡ 30-Second TL;DR

What Changed

M97 taped out, Q3 2024 launch with 700T+ compute for ADAS.

Why It Matters

Accelerates domestic AI chip adoption in EVs amid cost pressures, challenging import reliance. Positions Aixin as neutral supplier, fostering ecosystem for edge AI in homes/agents.

What To Do Next

Benchmark M97's DDR bandwidth in your ADAS simulator against Nvidia rivals.

Who should care:Enterprise & Security Teams

Key Points

•M97 taped out, Q3 2024 launch with 700T+ compute for ADAS.
•High DDR bandwidth (10k+ MHz) and leading process node maximize effective compute.
•Designed for VLA/world models, low power, cost-optimized architecture.
•H2 edge chips adapt mainstream LLMs like Qwen, support local AI Agents.
•Formed 'Qianli Alliance' with Qianli Tech, Jiyue Xingchen.

🧠 Deep Insight

AI-generated analysis for this event.

🔑 Enhanced Key Takeaways

•Aixin Yuanzhi (Aixin) has successfully transitioned from its initial focus on smart cockpit chips to high-performance ADAS/AD solutions, with the M97 representing its first major foray into the high-compute domain to compete with established players like NVIDIA and Horizon Robotics.
•The M97 architecture emphasizes 'effective compute' over theoretical peak TOPS, specifically targeting the memory wall bottleneck by utilizing advanced packaging and high-bandwidth memory interfaces to support the high data throughput required for end-to-end autonomous driving models.
•The 'Qianli Alliance' strategy is a deliberate ecosystem play aimed at reducing the integration friction for OEMs by providing a pre-validated hardware-software stack, effectively lowering the barrier to entry for deploying large-scale vision-language models (VLA) in mass-produced vehicles.

📊 Competitor Analysis▸ Show

Feature	Aixin M97	NVIDIA Orin-X	Horizon Journey 6
Effective Compute	700T+	254 TOPS	560 TOPS
Target Application	VLA/World Models	ADAS/AD	ADAS/AD
Market Positioning	Cost-optimized/High-efficiency	Premium/High-performance	Mass-market/High-efficiency

🛠️ Technical Deep Dive

Architecture: Utilizes a proprietary NPU architecture optimized for transformer-based models and VLA (Vision-Language-Action) workloads.
Memory Interface: Supports LPDDR5X-10666, addressing the memory bandwidth bottleneck common in high-compute ADAS chips.
Process Node: Manufactured on a leading-edge sub-7nm process node to optimize power efficiency (Performance-per-Watt).
Software Stack: Integrated support for mainstream LLMs (e.g., Qwen) and local AI Agent deployment, enabling on-device inference for cockpit-driving integrated systems.

🔮 Future ImplicationsAI analysis grounded in cited sources

Aixin will capture significant market share in the mid-to-high-end EV segment by 2027.

The focus on cost-optimized architecture for VLA models aligns with the industry trend of moving toward end-to-end autonomous driving in mass-market vehicles.

The Qianli Alliance will expand to include at least three additional Tier-1 suppliers by year-end 2026.

The alliance model is designed to scale rapidly by standardizing the hardware-software interface for OEMs.