Removing Obstacles before Breaking Through the Memory Wall - A Close Look at HBM Errors in the Field

Explore a groundbreaking study on High-bandwidth memory (HBM) errors in this 21-minute conference talk from USENIX ATC '24. Delve into the first systematic analysis of HBM errors, covering over 460 million error events collected from nineteen data centers over a two-year period. Discover how HBM's stacked architecture, while promising for overcoming the memory wall, introduces new reliability challenges. Learn about the unique error patterns exhibited by HBM compared to conventional DRAM, including differences in spatial locality, temporal correlation, and sensor metrics. Understand why traditional DRAM error prediction models fall short for HBM. Gain insights into Calchas, a novel hierarchical failure prediction framework designed specifically for HBM, which integrates spatial, temporal, and sensor information from various device levels. Examine the feasibility of failure prediction across hierarchical levels and the implications for future memory technologies.

USENIX ATC '24 - Removing Obstacles before Breaking Through the Memory Wall: A Close Look at HBM...