FinOps & Cost Intelligence
Dynamic Triton GPU/CPU auto-scaling, idle power-down, and cluster cost optimization
💰Cost Optimization Active
Total Cost Saved (YTD)
$14,523.48
GPU idle power-down savings
Active Inference Route
🟢 CPU FALLBACK
Zero GPU VRAM · Max cost savings
Cluster Efficiency
94.8%
Bin-packing ratio (KEDA optimal)
KEDA Scaling Policy
GPU Cluster Load0%
Inference RouteCPU Fallback
GPU StatusSUSPENDED (IDLE)
Active GPU Nodes0
GPU Threshold500 TPS
GPU powers on above this TPS
Policy
When TPS drops below 500, KEDA scales Triton to 0 pods and routes through XGBoost on CPU — saving ~$0.08/min in compute.
KEDA Auto-Scaler: Traffic TPS vs. Active GPU Nodes
When the spike ends (10:00), GPU nodes scale to 0 within 15s. XGBoost handles residual traffic on CPU — eliminating idle GPU cost entirely.