AI21 Deploys Kueue Scheduling on 10,000-GPU GKE Cluster, Cutting Hero-Job Starvation to 12 Hours

AI21 deploys Kueue automated scheduling across its roughly 10,000-GPU Google Kubernetes Engine cluster, cutting average hero-job starvation from 72 hours to 12 hours while Admission Fair Sharing prioritizes teams with fewer historical chip-hours.

• AI21 replaces manual GPU-allocation negotiations with Kueue automated scheduling across a shared GKE cluster of about 10,000 GPUs.
• The Kueue collaboration adds Admission Fair Sharing, which prioritizes teams with lower historical chip-hour use and addresses AI21's multi-node fairness gap.
• Topology Aware Scheduling cuts GPU fragmentation from 15% to 8% and reduces average hero-job starvation from 72 hours to 12 hours.