🤖 AI Summary
John Carmack has proposed an innovative concept to address the ongoing RAM crisis in AI by using fiber optic cables as a substitute for traditional RAM modules. He suggests that a long fiber cable, potentially up to 200 km, could function as a dynamic storage medium, continuously streaming data necessary for AI model inference and training. With demonstrated data rates of 256Tb/s over long distances, this approach could maintain a steady flow of data to AI processors, thereby reducing reliance on DRAM and its associated costs, while also promoting energy efficiency.
While Carmack's idea is still theoretical and presents challenges—such as maintaining signal integrity and the infrastructure needed—the concept could pave the way for future solutions in handling increasing memory demands in AI applications. He also discusses more immediate approaches, such as aggregating cheaper flash memory to enhance read bandwidth, contingent on the collaboration between flash and AI accelerator manufacturers. As the RAM shortage is expected to persist, exploring these alternatives may provide a sustainable pathway for AI advancements without the current constraints imposed by traditional memory technologies.
Loading comments...
login to comment
loading comments...
no comments yet