Eigen Exchange
AI workloads are driving rapid growth in data center energy demand, yet GPU clusters are often inefficiently utilized because capacity is reserved but unused, workloads are poorly packed, and fixed scheduling rules do not reflect changing priorities. Eigen Exchange is developing a market-based allocation layer for shared GPU infrastructure. It combines day-ahead allocation with real-time adjustments to match workloads to available GPUs while accounting for job size, timing, organizational priorities, and budgets. A discrete optimization engine is designed to improve packing for gang-scheduled jobs, while price signals encourage flexible workloads to shift away from constrained periods. Eigen is designed to integrate with Slurm and Kubernetes rather than replace existing schedulers. The project will evaluate impacts on utilization, queueing, stranded capacity, and the infrastructure and energy required per unit of computing through prototype development, workload simulations, and customer discovery.
Team Members
Jason Sun (MBA, GSB)