5.3 Cost levers¶
Everything you built today maps to a lever. Here's the ranking again with the exact knob you touched, so you can take it to your own cluster.
| # | Lever | Typical impact | The knob | Where you did it |
|---|---|---|---|---|
| 1 | Don't idle | 30–60% | consolidationPolicy: WhenEmpty + short consolidateAfter; KEDA minReplicaCount: 0 for dev; the < 5% for 30m alert |
5.2, 4.5 |
| 2 | Spot, diversified | 60–70% on the GPU line | karpenter.sh/capacity-type: [spot, on-demand] + ≥4 instance types + quota for both |
2.2 |
| 3 | Right-size | 20–75% per workload | time-slicing ConfigMap + node label; g6 (L4) instead of g5 (A10G) for inference; MIG on A100/H100 |
3.2 |
| 4 | Bin-pack | 10–30% | limits.nvidia.com/gpu on the NodePool; no gratuitous AZ spread; instance-gpu-count matched to the model |
2.2 |
| 5 | Pin AMIs | avoids one bad day per quarter | amiSelectorTerms: [{alias: al2023@v…}] |
2.2 |
| 6 | Shorten cold starts | lets you run leaner on 1 and 4 | pre-pull DaemonSet, EBS throughput, weights off the critical path | 3.1, 4.6 |
Three questions to take back¶
Answer them for your platform, not this one:
- What's the 30-day average
DCGM_FI_DEV_GPU_UTILacross your fleet? If you don't know, that's the first dashboard. If it's under 40%, lever 1 and 3 pay for a quarter of platform work in a month. - What happens when a Spot GPU node is reclaimed at 3 a.m.? If the answer involves a human, you need a second replica, a PDB, and
terminationGracePeriodSeconds— before you turn on Spot, not after. - Who owns the
forgotten-notebooks? If GPU pods don't carry ateamlabel, nobody does. Admission policy (Kyverno/Gatekeeper) that requires the label on any pod requestingnvidia.com/gpuis a one-afternoon change.
Cost of this workshop¶
For the record, roughly what your account will show for today, all On-Demand worst-case:
| Hours | $ | |
|---|---|---|
| Base stack (overnight + session) | ~16 | ~9 |
| GPU node A | ~2.5 | ~2.5 |
| GPU node B (the notebook) | ~1.2 | ~1.2 |
| GPU node C (the scale-out) | ~0.4 | ~0.4 |
| NAT data (3 image pulls, weights) | ~1.5 | |
| Total | ~$15 |
…provided you run Clean up before you close the laptop. Which is the last thing we do together.