Skip to main content
Rightsizing · High severity

GPU Instance Low Utilisation

An accelerator is billed whole and by the hour whether or not the model keeps it busy, and it is the most expensive unit of idle capacity available to rent.

Free to start. No card. The playground just needs your work email.

AWSprovider
Highseverity
Savingswhere it shows up
Manualfix, steps written out

AWS

Found, explained, handed over.

Findings with a dollar figure attached, idle, oversized, orphaned, unscheduled and undiscounted spend.

01

Detect

ZopNight checks this automatically across AWS, with read-only access to the account.

02

Explain

Every finding says exactly what to change: Rightsize the underutilised GPU instance. Where the saving can be proven it is priced; where it cannot, the finding says so.

03

Fix

The finding opens with the fix already written out, step by step, so it is one ticket, not an investigation.

Applies to
EC2 Instances on AWS
The fix, by hand
  1. CloudWatch → Metrics → /aws/sagemaker or EC2 GPU metrics → GPUUtilization on the resource. Confirm sustained low utilisation across 30 days.
  2. Decide: downsize (e.g. p3.2xlarge → g5.xlarge), schedule (Stop overnight for dev/training), or move to Spot for fault-tolerant training.
  3. For downsize: validate ML workload on the smaller GPU in staging first (memory + VRAM matter as much as compute).
  4. EC2 → select the resource → Stop → Change instance type → start.
  5. For training fleets: switch the job to ManagedSpotTraining to capture Spot discounts with automatic checkpoint resume.
  6. Monitor GPU + step-time metrics for 48h post-change.

These are the steps the finding carries in the product.

Category
Rightsizing. Resources sized for headroom they never use. A smaller size runs the same workload for less.
Where it appears
The Savings tab of Recommendations, with every affected resource listed.
Rule ID
RC-1500

Related checks

See the oversized resources in your account.

Connect a read-only role and the first pass runs on your own estate. This check, and the rest of the catalogue, with it.

Prefer to talk it through first? Book 20 minutes with the team.

  • $30M+annualised cloud spend under management
  • 550K+resources tracked since launch
  • 20-60%off the bill in the first month
  • SOC 2Type II report, plus ISO 27001

Figures published on zop.dev.

Multi-cloud automation· Production-ready in 30 min· SOC 2 · ISO 27001· 20–60% off the bill, first month· 4 platforms · 1 console· Multi-cloud automation· Production-ready in 30 min· SOC 2 · ISO 27001· 20–60% off the bill, first month· 4 platforms · 1 console·