原创 理想TOP2与群友 2025-09-05 12:42 四川
This is Yan. Let me share some information about our in-house design chip. The chip successfully tapered out and returned at the beginning of this year. And it is currently ongoing in vehicle testing and everything is in good shape. We expect to deploy it on our flagship models and deliver it to users next year. It takes about 3 years from setting up the project and to its shipment. As far as I know, it is the fastest among similar efforts. The performance is quite satisfactory, compared to the most powerful chips on the market, it could provide 2x performance when running GPT like large language models and 3x when running vision models like CNN. We designed a novel data flow architecture in which model competition is mostly driven by data, not instructions like other architectures.
And in this way, the chip could achieve higher parallelism at the runtime, and we believe it is more suitable for large neural networks. The data-driven logic is orchestrated by our in-house design compiler allowing the hardware to be more efficient and running at a higher frequency than most comparables in the market. Quite different from other AI chips on the market, we adopted a truly — a true hardware software co-design approach, the chip, the compiler, the runtime system and the halo operating system are designed together from the beginning. So we can vertically integrate the hardware and software modules to a more powerful AI inference system and more easily, and it could keep on scaling in the future. With the landing of VLA Models on vehicles, we observed that computing power increase could translate to ADAS performance increase better than before, which means higher the computing power, the better the performance, and it’s more predictable.
We have very strong confidence our innovative architecture as well as the full stack development capability could become our continuous differentiated capabilities and grow even stronger in the future.
来源:Insider Monkey链接:https://www.insidermonkey.com/blog/li-auto-inc-nasdaqli-q2-2025-earnings-call-transcript-1599680/中文翻译:大家好,我是谢炎。我来分享一些关于我们自研芯片的信息。这款芯片在今年年初已成功完成流片并返回。目前,我们正在进行车辆测试,一切进展顺利。我们预计明年将这款芯片部署在我们的旗舰车型上,并交付给用户。从项目立案到最终交付上车,整个过程大约需要3年时间。据我所知,这在同类项目中是速度最快的。芯片的性能表现相当令人满意。与市场上最顶尖的芯片相比,在运行像GPT这样的大语言模型时,它的性能可以达到前者的2倍;在运行像CNN这样的视觉模型时,性能可以达到3倍。我们设计了一种新颖的数据流架构(data flow architecture)。在这种架构中,模型的计算主要由数据驱动,而非像其他架构那样由指令驱动。通过这种方式,芯片在运行时可以实现更高的并行度,我们相信它更适合大型神经网络。这种数据驱动的逻辑由我们自研的编译器进行调度(orchestrated),使得硬件能够更高效地运行,并且其运行频率也高于市场上大多数的同类产品。与市场上的其他AI芯片截然不同,我们采用了一种真正的软硬件协同设计(hardware software co-design)方法。芯片、编译器、运行时系统(runtime system)以及我们的操作系统,从一开始就是作为一个整体共同设计的。因此,我们可以更轻松地将软硬件模块进行垂直整合,从而打造出一个更强大的AI推理系统,并且这个系统在未来能够持续扩展升级。随着VLA模型在车端的落地,我们观察到,算力的提升比以往任何时候都更能有效地转化为高级驾驶辅助系统(ADAS)性能的提升。这意味着算力越高,性能就越好,并且这种正向关系也变得更具可预测性。我们坚信,我们创新的架构以及全栈自研的能力,将成为我们持续的、差异化的核心竞争力,并将在未来发展得更加强大。加微信,进群深度交流理想实际经营情况与长期基本面。不是技术群,车友群。
