Coordination and Workload Algorithms to Unlock the Potential of Polymorphic Systems of Systems

Explore advanced coordination and workload algorithms designed to optimize AI deployment across heterogeneous computing architectures in this 20-minute conference talk from the Open Compute Project. Learn how the rapidly evolving AI landscape requires sophisticated orchestration of disaggregated, composable, and heterogeneous resources to handle large-scale AI workloads effectively. Discover the AI HW-SW CoDesign workstream's algorithmic solutions, including graph sharding algorithms that work with emulation software to predict optimal deployment strategies for established workloads. Examine algorithmic priors that transform workloads into more efficiently distributable compute graphs, enabling AI workloads to co-evolve with emerging heterogeneous hardware stacks. Understand how these coordination algorithms address two critical questions: how existing workloads can be best coordinated across heterogeneous architectures, and which new workloads become scalable through heterogeneous compute. Gain insights into creating new AI systems that would not be scalable on current technology by leveraging polymorphic compute stacks with diverse accelerators, fabrics, and storage solutions.