
Proposed using a long fiber loop as an L2 cache to hold AI model weights for near-zero latency and gigantic bandwidth.
How media typically covers John Carmack
Referenced in coverage
John Carmack proposes using long-distance fiber optic loops as an L2 cache for AI model weights, offering potential power efficiency advantages over DRAM while maintaining gigantic bandwidth and near-zero latency.
“Proposed using a long fiber loop as an L2 cache to hold AI model weights for near-zero latency and gigantic bandwidth.”