This clip provides an overview of composable disaggregated systems and infrastructures and highlights the challenges faced in resource management within heterogeneous infrastructures.
Modern large-scale workloads, such as machine learning, require significant computing resources. However, traditional compute clusters are not efficient for these workloads as they over-provision resources, resulting in wasted energy, storage and cost. This clip discusses the concept of composable disaggregated infrastructures (CDI), which physically disaggregate computational resources over high-speed and low-latency fabrics. CDI allows for dynamic composition of resources based on workload demands, eliminating the need for overprovisioning. This approach enables real-time allocation of resources, efficient platforms, and reduced electricity costs. While network disaggregation is already prevalent for storage devices, the trend is extending to computational engines, memory elements, and accelerators, catering to modern and future high-performance computing applications.