Operations · Advanced
GPU compute efficiency auditor
Evaluate true hardware throughput and detect underutilized accelerator clusters
- On a scheduleEvery six hours
- ThenSample throughput and memory metrics
- ResultReview efficiency summary in Slack
Standard GPU metrics often show full utilization while actual compute cores remain idle. This bot inspects hardware performance counters across compute instances, contrasts active kernel runtime with real throughput metrics, and compiles diagnostic efficiency digests to prevent overprovisioning.
Prompt to start with
BotBentoRecommended
Run diagnostic profiling commands on target accelerator nodes via the terminal. Query hardware performance metrics to compare memory bandwidth and arithmetic throughput against reported kernel active times. Format findings into a cluster efficiency digest highlighting bottlenecks or idle capacity. Before posting any report to engineering channels or making infrastructure alert modifications, prompt the engineer for explicit confirmation.