用光纤充当缓存？芯片被颠覆了

Core Viewpoint - The article discusses the potential of using fiber optics instead of silicon to define how artificial intelligence stores and retrieves knowledge, as proposed by John Carmack, highlighting the advantages of fiber optics in data transmission and storage [2][3]. Group 1: Fiber Optics in AI - John Carmack suggests using fiber optic loops as high-speed data caches for AI models, which could revolutionize data storage and retrieval methods [2]. - Current single-mode fiber can transmit data at a speed of 256 terabits per second over 200 kilometers, allowing for approximately 32 GB of information to be stored at any given moment [2]. - The proposed method would function as a secondary cache, enabling model weights to be stored at light speed with minimal latency and significantly higher bandwidth compared to traditional memory [2][3]. Group 2: Advantages Over Traditional Memory - Fiber optics offer predictable performance, low power consumption, and substantial bandwidth potential compared to volatile DRAM, which requires constant refreshing of electrical signals [3]. - The efficiency advantages of fiber optics are appealing, especially as the miniaturization of components slows down, potentially making fiber optics a more favorable option than DRAM [3]. Group 3: Challenges and Future Directions - A significant challenge is the high cost of high-quality fiber optics over long distances, which may offset energy savings from reduced power consumption [3]. - Carmack's more practical next step involves tightly coupling flash memory chips with AI accelerators to facilitate rapid movement of model weights without relying on DRAM, requiring collaboration between semiconductor manufacturers and accelerator designers [4]. - Research teams are exploring architectures that utilize solid-state storage, indicating a trend towards blurring the lines between storage and memory [4].