AI Memory Stored in 1,000 km of Fiber Optic Cable by Uppsala University Researchers
Researchers at Uppsala University are developing an innovative method to store artificial intelligence (AI) memory within light signals transmitted through 1,000 kilometers of optical fiber. Their project, dubbed 'Fiber Memory,' involves circulating the data of a large language model (LLM) in a continuous loop within the fiber optic network. This approach aims to revolutionize how AI models are accessed and utilized, particularly in large data centers. Current AI data centers, housing approximately 10,000 computing chips, typically store a complete copy of the AI model on each chip's onboard memory, known as High Bandwidth Memory (HBM). For instance, a compressed Llama 3 70B model requires 70 GB of storage. Storing this model 10,000 times across all chips results in an exorbitant memory requirement of 700 terabytes, with significant redundancy. The Fiber Memory project seeks to overcome this inefficiency by centralizing the model's memory in the optical fiber infrastructure, potentially reducing costs and computational overhead.
This research proposes a novel paradigm for AI memory management, shifting from distributed, redundant storage on individual processing units to a centralized, light-based system within optical fiber. By circulating LLM data through extensive fiber optic networks, the project addresses the significant cost and energy inefficiencies associated with replicating massive models across thousands of specialized chips. This approach could fundamentally alter data center architecture, reducing the need for vast amounts of expensive onboard memory. Future implications may include more scalable and sustainable AI deployments, though challenges related to data access latency, signal integrity over long distances, and integration with existing hardware will need to be thoroughly addressed. The long-term viability hinges on demonstrating comparable or superior performance to current methods while achieving substantial cost and energy savings.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.