
Opis
NVIDIA GPU infrastructure spans hardware, system software, networking, storage, orchestration, MLOps, and inference. Understanding how these components fit together, where their responsibilities overlap, and which distinctions matter requires a clear, structured path.This book provides that path through one coherent narrative of the NVIDIA GPU infrastructure stack. It covers accelerated computing, CUDA, and the NVIDIA software ecosystem before comparing data center GPUs against workload characteristics. You will examine MIG, vGPU, and DCGM for resource sharing and monitoring; Kubernetes and Slurm for GPU scheduling; and the networking and storage layer, including Ethernet, InfiniBand, RDMA, GPUDirect Storage, BlueField DPUs, and DOCA.
Later chapters connect infrastructure to the AI lifecycle through Airflow, MLflow, and Kubeflow for MLOps, NGC for software delivery, and ONNX, TensorRT, and Triton for inference. You will also explore the Kubernetes components, monitoring technologies, scaling considerations, and diagnostic concepts that support production GPU clusters.
By connecting these technologies instead of presenting them as isolated products, the book helps you compare platform choices, understand component boundaries, discuss trade-offs, and develop a durable mental model of NVIDIA GPU infrastructure.