ECE Seminar - Journey through Computer Engineering: Math/Algorithms–languages–frameworks chips–systems 计算机工程之旅：数学/算法 → 语言 → 框架 → 芯片 → 系统

HiSilicon’s AI chips are not only about transistors or TOPS. They are about making intelligence usable. We moved compute out of the datacenter and into daily life. That shift—bold, sometimes messy—made intelligence more accessible. In this talk, we open the hood and share what actually worked when building widely deployed AI silicon.

The next step is tougher. We are moving from “on-device AI” to embodied systems that perceive, reason, and act in the real world. Getting there means tight co-design across sensing, processing, and actuation, all under strict power limits. It also calls for math we can implement—ideas that become efficient algorithms—plus memory and state that let machines adapt over time.

Our lens is practical: “show me the code.” We link programming languages and frameworks to real hardware choices and open collaboration. Beyond AI, we discuss what this implies for general-purpose computing: memory, processing, and chip design must move together if we want better performance and efficiency.

Finally, we outline a near-term plan toward a more capable chip: event-based sensing, on-chip lifelong learning, and bio-inspired blocks. Academia is the right crucible for these high-risk ideas. Let’s identify the missing pieces—and build systems where intelligence is not only powerful, but also embodied.

海思的 AI 芯片，不只是晶体管数量或 TOPS 指标；它关乎让智能可用。我们把计算从机房推到日常设备里。这个转变很大胆，也不总顺利，但确实让智能更可达。本次报告将“打开引擎盖”，分享在大规模落地 AI 芯片过程中真正有效的做法与经验。

下一步更难：从“端上 AI”迈向具身系统——能在真实世界中感知、推理与行动。要做到这一点，需要在感知—处理—执行之间进行紧密协同设计，并在严苛功耗下完成工程落地。同时，我们需要能实现的数学，把想法落实为高效算法；还需要让系统拥有记忆与状态，得以随时间自适应。

我们的视角是务实的：“show me the code”。我们把编程语言、框架与硬件抉择和开放协作对应起来。超越 AI 话题，这也指向通用计算的下一步：存储、处理与芯片架构必须同步前进，性能与能效才可能同时提升。

最后，我们给出一条近期可执行的路线：事件驱动感知、片上终身学习与类脑/仿生结构模块。高校与实验室是孕育高风险创新的最佳土壤。让我们一起补齐关键拼图，打造既强大、也扎根于物理世界的智能系统。