Li Auto Mach 100 Paper Accepted to ISCA

Post LinkedIn

🏠Read original on IT之家

#dataflow-arch #ai-chip #automotive-aimach-100

💡Auto 1st ISCA paper: AI dataflow chip 3x Nvidia Thor effective perf.

⚡ 30-Second TL;DR

What Changed

Mach 100 paper accepted to elite ISCA 2026 Industry Track; auto first

Why It Matters

Validates custom dataflow chips for AI in autos, challenging Nvidia in edge inference. Signals industry shift to efficient, evolvable architectures.

What To Do Next

Track the 2026 ISCA Mach 100 paper for dataflow insights in AI chip design.

Who should care:Researchers & Academics

🧠 Deep Insight

Web-grounded analysis with 1 cited sources.

🔑 Enhanced Key Takeaways

•The Mach 100 chip is a core component of Li Auto's 'software-hardware co-design' strategy, specifically engineered to eliminate the data-shuttling bottlenecks inherent in traditional GPGPU architectures when running large 3D Vision Transformer (ViT) models.
•Beyond the chip, the new L9 Livis model integrates the Mach 100 with a 'full-form' by-wire chassis, featuring steer-by-wire, four-wheel steering, and the first production-ready fully electronic mechanical braking (EMB) system compliant with new national standards.
•Li Auto has simultaneously launched 'MindVLA-o1', a next-generation autonomous driving foundation model designed to be hardware-agnostic, supporting the dual-Mach 100 configuration, Nvidia Thor-U, and legacy dual-Orin X platforms.

📊 Competitor Analysis▸ Show

Feature	Li Auto L9 Livis (Mach 100)	XPeng GX (Turing AI)	Nvidia Thor-U (Reference)
Compute/Chip	1280 TOPS (Effective)	N/A	700 TOPS (Rated)
Total Compute	2560 TOPS (Dual)	3000 TOPS (Quad)	N/A
Architecture	Dataflow (Custom)	N/A	GPGPU (Standard)
Key Differentiator	Software-Hardware Co-design	High-volume integration	Industry standard platform

🛠️ Technical Deep Dive

Architecture: Dataflow-based design optimized for direct compute-unit data transfer, bypassing traditional memory-shuttling overhead.
Process Node: 5nm automotive-grade manufacturing.
Performance: 1280 TOPS effective compute per chip; 2560 TOPS total in dual-chip configuration.
Model Optimization: Specifically tuned for VLA (Vision-Language-Action) large models and 3D ViT (Vision Transformer) workloads.
System Integration: Part of a closed-loop stack including StarRing OS and the MindVLA-o1 foundation model.

🔮 Future ImplicationsAI analysis grounded in cited sources

Li Auto will transition its entire L-series lineup to the Mach 100 architecture by end of 2026.

Company executives have publicly stated that the L9 refresh serves as the primary deployment platform and that lessons learned will trickle down to standardize 800V and advanced computing across the L-series.

The Mach 100 architecture will enable L3/L4 autonomous driving capabilities in production vehicles.

The combination of 2560 TOPS, 360-degree lidar coverage, and a full by-wire chassis provides the necessary redundancy and compute overhead for higher-level autonomy.