DIMSUM: Diving into Multihost Shared and Unified Memory

DIMSUM: Diving into Multihost Shared and Unified Memory

Learn about multi-host shared memory devices and their potential impact on modern AI through this technical presentation. Explore how disaggregation of memory into large shared pools can reduce total memory costs and enable efficient memory sharing while minimizing internode communication overhead. Discover an optimal design for shared memory devices and follow along with a practical AI workload implementation using Anyscale Ray. Understand the necessary software transformations required to leverage shared memory devices effectively in datacenter environments, as explained by Staff Software Engineer Jon Hermes from Arm.