'A disk in a planet-scale computer': Meta has so many expensive GPUs that it's buying SSDs to kill idle time
Meta rebuilt storage systems after slow data repeatedly stalled expensive AI GPUs SSD caching dramatically reduced AI dataset loading times from hours to minutes Meta replaced complex metadata lookups with a faster unified storage architecture Meta says storage systems have failed to keep pace with AI computing power, creating delays that leave costly GPUs waiting instead of processing workloads efficiently. According to the company's engineers, storage bottlenecks remain a major cause of GPU stalls, increasing operating costs while slowing research progress and extending development timelines. To address those delays, Meta redesigned its storage architecture, arguing that faster movement of data can unlock greater value from expensive AI hardware investments. Meta rebuilds storage architecture to keep GPUs working Meta's engineers explained that the company operates hundreds of exabyte-...