Physical AI 如何扩展数据移动架构
Scaling Data Movement For Physical AI
Physical AI 系统混合多种计算单元，需实时处理数据、做出决策并兼顾电池续航，关键在于决定数据在哪里处理以及如何最小化数据移动。Arteris 产品管理及营销总监 Ashley Stevens 解释，一致性互连更适合 CPU，非一致性互连更适合 AI 加速，并展望了工程师后续面临的其他挑战。
One size does not fit all.
Physical AI poses a host of new challenges for chip engineers. These systems contain a mix of compute elements, which collectively must process a stream of real-time data, analyze it, make decisions, and still function long enough on a battery to be useful. The key here is deciding what data gets processed where, and how to minimize the movement of data. Ashley Stevens, director of product management and marketing at Arteris, talks explains why data movement plays such an important role in these systems, why coherent interconnects are better suited to CPUs and non-coherent interconnects are better for AI acceleration, and what other challenges engineers will face going forward.
Ed Sperling
(all posts)
Ed Sperling is the editor in chief of Semiconductor Engineering.
来源： Semiconductor Engineering 芯片与封装 · semiengineering.com