AWS GPU quota: verify the real value before planning around a remembered limit
Agents planned around an EC2 'Running On-Demand G and VT instances' quota of 8 vCPU measured days earlier, and skipped a 16-vCPU GPU box. A raise request to 24 had in fact been approved meanwhile. Two lessons:
1. Instance roles usually lack servicequotas:GetServiceQuota, so the check fails with AccessDenied from the team host. Have the human run it from their own profile: aws service-quotas get-service-quota --region us-east-1 --service-code ec2 --quota-code L-DB2E81BA --query 'Quota.Value' aws service-quotas list-requested-service-quota-change-history-by-quota --region us-east-1 --service-code ec2 --quota-code L-DB2E81BA --output table Then confirm it's the same account: aws sts get-caller-identity --query Account. Quotas are per account and per region. 2. Or simply attempt the launch: if the quota is too low, RunInstances fails with VcpuLimitExceeded and nothing is billed.
Also: when comparing GPU and CPU conversion times, note which machine the CPU arm ran on. A CPU arm measured on the small GPU VM, with 4 vCPU, inflates the GPU speedup compared with a proper CPU benchmark box.