Recently, BYD has entered the final phase of depleting its non-intelligent vehicle inventory, soon to initiate equal rights for intelligent driving. The conference had some highlights, especially Bosch’s speech on intelligent driving for traditional vehicles. Today, let’s briefly analyze the intelligent driving chip architectures of two third-party companies.
1. Black Sesame Intelligence: Self-developed NPU Architecture with Vertical Integration
Features of the “Jiushao” NPU Architecture
Based on the self-developed “Jiushao” NPU architecture, it employs a big core design that supports real-time inference of large Transformer models, with energy efficiency significantly superior to traditional solutions. This architecture optimizes the data flow path to avoid computational bottlenecks, making it suitable for rapid iteration of end-to-end intelligent driving algorithms.
Safety and Determinism: The industry’s highest safety level NPU design can avoid random errors during model inference, ensuring consistency between training and deployment, meeting automotive-grade high reliability requirements.
Multi-task Integration: A single chip integrates CPU, GPU, NPU, ISP, and other modules, supporting cross-domain integration of intelligent driving and cabin functions (e.g., Wudang series chips).
Computing Power and Cost Advantages
The A2000 series chips cover computing power from 256 to 1000 TOPS, which is more scalable compared to NVIDIA’s Orin-X (254 TOPS), while the unit price is as low as $15 per chip (NVIDIA Orin-X is about $500 per chip), offering both high cost-performance and low power consumption.
2. Horizon: BPU-driven Soft-Hard Collaborative Strategy
Features of the “Zhengcheng” Series BPU Architecture
Adopting the concept of “scene-defined chips,” it achieves high energy efficiency through soft-hard collaboration optimization. The Zhengcheng 5 chip has a computing power of 128 TOPS, which, although lower than Black Sesame’s A1000 Pro (196 TOPS), performs better in low-power scenarios due to deep adaptation of algorithms and hardware.
Flexibility: Supports multi-sensor fusion and rapid algorithm iteration, adapting to L3+ high-level autonomous driving needs, such as the Horizon Superdrive technology enhancement scheme.
Ecological Cooperation: An open algorithm licensing model, jointly developing customized solutions with automakers (e.g., Li Auto, Changan, etc.), forming a broad customer network.
Market Performance and Commercialization Capability
By the end of 2023, Horizon’s market share in China’s high-level intelligent driving market reached 35.5% (NVIDIA at 50.8%), covering 285 vehicle models with a high repurchase rate; Black Sesame’s market share is 7.2%, focusing on the mid-to-low-end market.
International Layout: Established a joint venture with Volkswagen to accelerate globalization, while Black Sesame currently focuses on domestic automakers (e.g., FAW, SAIC).
3. Core Differences Summary
| Dimension | Black Sesame Intelligence (NPU) | Horizon (BPU) |
|---|---|---|
| Technical Focus | Self-developed hardware architecture + cross-domain integration | Soft-hard collaboration + scene-driven |
| Computing Power Performance | High computing power (up to 1000 TOPS) | Medium to high computing power (Zhengcheng 6 is 10-560 TOPS) |
| Commercialization Path | High cost-performance entry into mid-to-low-end market | Ecological cooperation dominates high-level intelligent driving market |
| Safety Certification | First in the country to pass ISO 26262 ASIL-B certification (2021) | Completed the same level certification in 2022 |
4. Future Competition Focus
Both parties need to break through in large model adaptability (e.g., Transformer support), cross-domain integration capability (integrated cabin and driving), and global cooperation. Black Sesame needs to accelerate penetration into the high-end market, while Horizon faces differentiation challenges under NVIDIA’s computing power pressure.
The Horizon robot will be fully released on April 24, and it is expected to sell well; hopefully, it will have a good price.