智驾芯片算法专家交流 20250807 摘要 华为新一代芯片将提供 500-800 TOPS 算力版本，采用单芯片方案替代 双芯片，解决特征层传输局限性，并降低成本，预计价格在一万多美元， 低于双芯片方案。 车端芯片架构基于达芬奇架构，偏向整形运算，与云端服务器的浮点运 算需求不同，导致成本差异显著。华为未转向 GPGPU 方向，而是优化 ASIC 架构，推进存算一体化，提高数据吞吐效率。 华为自动驾驶算法正从 IDS3.1/3.3 的两段式结构向端云协同 Vivo 框架 转变，通过云侧世界引擎模型生成训练数据，蒸馏出 MOE 多专家原生 基模型，提高对复杂场景的泛化能力。 当前多模态大语言模型参数量约 1.几个 B，低于特斯拉和理想，车端硬 件受限于算力和带宽，运行 1.5B-2B 模型需 40GB/s 带宽。数据质量对 训练效果至关重要，高质量数据标注和工程是提升体验的关键。 华为车端多模态大语言框架可实现 100 毫秒内出结果，全链路 200 毫 秒，融合感知、后处理、预测、规划及控制，通过注意力机制提高效率。 系统基于盘古大模型，并结合开源资源进行自主创新。 Q&A 下一代 MDG1,000 芯片在算 ...

Summary of the Conference Call Company and Industry Overview - **Company**: 和胜新材 (He Sheng New Materials) - **Industry**: Artificial Intelligence and High-Performance Composite Materials - **Subsidiaries**: 海希技术 (Hai Xi Technology) and 易之电子 (Yi Zhi Electronics) Key Points and Arguments 1. **Investment in AI**: 和胜新材 has invested in 海希技术 to enter the AI sector, focusing on domestic AI chips and providing servers and AI integrated machines for intelligent computing centers. Clients include major internet companies like ByteDance and Tencent [2][3] 2. **Market Demand**: The market shows significant demand for domestic AI integrated machines, with the price for the DeepThink 六七 1 full version hardware deployment ranging from 2.5 million to 5 million yuan [2][5] 3. **Strategic Partnerships**: 海希技术 has a two-year agreement with 图耀 (Tu Yao) to deliver 800 AI integrated machines, having already procured 80 units [2][5] 4. **Investment in 易之电子**: 和胜新材's investment in 易之电子 aims to reduce costs and enhance market competitiveness, focusing on ARM and RISC-V architecture chips [2][6] 5. **Product Performance**: 易之电子's TF7,000 chip, based on ARM A8 architecture and 7nm process, is favored for its cost-effectiveness and low energy consumption, already supplying to Tencent Games and undergoing tests with ByteDance [2][7][9] 6. **Domestic Replacement Progress**: The pace of domestic replacement has accelerated, with 海希技术 adapting to most suppliers, including R1 and V3 versions, showing faster-than-expected progress [2][11] 7. **AI Model Development**: The DeepSeek model has driven the development of domestic large models, with applications rapidly expanding in downstream industries [2][5] 8. **Cost Sensitivity**: The adoption of integrated machines by internet giants is influenced by cost sensitivity, with 和胜新材's DeepSeek model offering a competitive price advantage [2][10] 9. **Energy Efficiency**: The Manycore 64 deep neural network accelerator box developed by 易之电子 shows energy efficiency advantages, particularly in lightweight models, achieving over 83% efficiency compared to mainstream GPUs like NVIDIA's A100 and H100 [2][12] 10. **Adaptation to Domestic Ecosystem**: The GPU from 易之电子 demonstrates high adaptation efficiency with the Deepseek21 large model, supporting the domestic ecosystem migration [2][13][14] Additional Important Information - **Company Background**: 和胜新材, established in 2002 and listed in 2009, is a leading supplier of high-performance composite materials, focusing on innovative material research and market expansion [3] - **Future Outlook**: The company is optimistic about the overall market scale and growth potential in both traditional and AI-related sectors, supported by favorable industry policies [3][10]