hard20 pts · first answer countsgpu-performancemonitoring

GPU shows 100% but is underused

nvidia-smi shows 100% GPU utilisation for an inference server, yet profiling shows low arithmetic throughput and small batches. How can both be true?

A.nvidia-smi measures GPU temperature, not useB.The GPU is brokenC.The utilisation number shows the share of time any kernel was running, not how much of the GPU compute units were busyD.Inference cannot use the GPU fully by design

← Semantic cache returns wrong answers High drift, stable performance →