Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore a comprehensive conference talk examining the critical challenge of memory underutilization in data centers, where over half of provisioned DRAM remains unused despite representing one of the largest cost components in server infrastructure. Learn how major cloud providers like Microsoft Azure, Google Cloud, and Meta regularly experience memory utilization rates of 50% or below, resulting in massive wasted spending across the industry's $100+ billion annual DRAM expenditure. Discover the limitations of previous industry attempts to address this challenge through technologies like CXL and Optane, and understand why these approaches have fallen short of solving the fundamental problem. Examine MEXT's innovative AI-powered predictive memory solution that enables Flash storage to operate at DRAM-speed performance by intelligently offloading underutilized memory pages to Flash while using neural network-based AI to predict which pages should be preloaded back into DRAM. Understand how this approach maintains application performance within significantly smaller DRAM footprints, delivering substantial cost reductions across any computing environment including on-premises, cloud, virtualized, bare metal, and containerized deployments without requiring changes to operating systems or applications. Gain insights into the technical implementation of AI techniques adapted from natural language processing to predict memory access patterns rather than words, and explore the potential for unprecedented cost efficiencies and capacity expansion in modern computing infrastructure.
