After installing drivers on an NVIDIA Tesla A100, the GPU reports a constant 25% utilization despite no workload, which can be resolved by enabling persistence mode using a simple nvidia‑smi command to keep the driver loaded and improve performance stability.
When deploying a Kubernetes cluster with kubespray and the NVIDIA runtime, Containerd 2.x reports "nvidia‑smi not found" because the go‑toml v2 parser treats the "binaryName" key differently, causing the wrong runtime wrapper to be used; the article details the configuration inspection, version comparison, code demonstrations, and practical work‑arounds.
The article explains why nvidia-smi cannot display GPU processes when run inside a container, analyzes the underlying pid‑namespace isolation and kernel‑level restrictions, and provides three practical solutions—including using hostPid, custom kernel interception modules, and the nvitop tool—plus a workaround for gpu‑operator deployments.
The article investigates why Nvidia L40S and RTX A6000 GPUs show less memory via nvidia‑smi than their advertised 48 GB, revealing that enabled ECC memory reserves a few gigabytes, and demonstrates the effect by toggling ECC on a Tesla‑T4 card.
This article explains Nvidia's Multi‑Instance GPU (MIG) technology, compares it with vGPU, walks through enabling and partitioning MIG on A100 cards using nvidia‑smi commands, and shows how to expose MIG resources in Kubernetes with single and mixed strategies.
