Why We Moved Our Timefold Java Worker Pods from AMD to ARM64

Timefold engineers discovered that AMD x86-64 cloud instances with hyperthreading caused severe performance variance in their Java-based optimization solver pods. When two solver jobs landed on the same node, they competed for shared physical cores, causing move calculation speed to drop by more than half. Switching to ARM64 instances (GCP Tau/Ampere Altra or AWS Graviton), where each vCPU maps to a physical core with no hyperthreading, eliminated the variance and allowed 16 concurrent jobs to scale gracefully. The ARM instances also proved roughly twice as cost-effective because AMD's effective utilization under concurrent compute-intensive load was far lower than the vCPU count suggested.