دانلود کتاب اصول زیرساخت پردازنده گرافیکی انویدیا، راهنمای ساختاریافته برای زیرساخت پردازنده گرافیکی انویدیا، از CUDA تا عملیات تولید

AI infrastructure is no longer defined by the GPU alone. Modern accelerated platforms combine GPU hardware with system software, high-speed networking, storage, orchestration, monitoring, virtualization, MLOps, and inference technologies. Each layer solves a different part of the problem, but the real challenge for infrastructure professionals is understanding how those layers interact and where the responsibility of one technology ends and another begins. The NVIDIA ecosystem reflects this breadth. CUDA provides the foundation for GPUaccelerated computing, while technologies such as Tensor Cores, NVLink, and NVSwitch influence how workloads execute and scale. MIG and vGPU provide different approaches to sharing accelerator resources, DCGM supports GPU monitoring and management, and Kubernetes and Slurm provide different scheduling models. Beyond the GPU itself, technologies such as InfiniBand, RDMA, GPUDirect Storage, BlueField DPUs, and DOCA shape how data moves through the platform. At the application and operations layers, tools such as NGC, Airflow, MLflow, Kubeflow, ONNX, TensorRT, and Triton help move AI workloads from development into production. With so many technologies involved, learning them as a catalogue of individual products is rarely enough. An infrastructure decision usually depends on the workload and the boundary at which a problem occurs. A requirement for hardware-isolated GPU resources is different from a requirement for GPU-enabled virtual machines. A model that is starved for input data needs a different response from one that is compute-bound. A container that cannot access a GPU presents a different problem from a pod that cannot be scheduled or an inference service that has poor throughput. Understanding these distinctions is what turns product knowledge into useful infrastructure judgment. This book presents NVIDIA GPU infrastructure as one connected system. It begins with the foundations of AI workloads and accelerated computing before introducing CUDA and the NVIDIA software stack. You will then examine the architecture and characteristics of NVIDIA data center GPUs and learn how workload requirements influence platform selection. From there, the book moves into the operational layers surrounding the accelerator, including GPU partitioning, monitoring, scheduling, storage, networking, virtualization, and DPU-based infrastructure offload. The later chapters connect this infrastructure to the AI lifecycle. You will see how Airflow, MLflow, and Kubeflow support reproducible machine learning workflows; how ONNX, TensorRT, and Triton serve different roles in production inference; and how NGC, Kubernetes, the NVIDIA Container Toolkit, GPU Operator, and monitoring technologies combine into an operable GPU platform. Troubleshooting scenarios throughout the book reinforce the idea that infrastructure should be investigated as a set of connected layers rather than as isolated components. The aim is not to make you memorize every NVIDIA product or specification. GPU generations, software versions, and platform capabilities will continue to evolve. Instead, this book focuses on the more durable knowledge: what each technology is designed to do, how it relates to the surrounding stack, what trade-offs distinguish similar technologies, and how to reason from a workload requirement or operational symptom to an appropriate infrastructure response. By the end of the book, you should have a structured mental model that allows you to discuss, evaluate, and work with NVIDIA GPU infrastructure with greater confidence. This book is for system administrators, cloud and DevOps professionals, data center and networking teams, solution architects, technical managers, presales professionals, and anyone moving into AI infrastructure roles who needs a structured understanding of the NVIDIA GPU ecosystem. It is also suitable for readers who work alongside AI, MLOps, or platform teams and want to understand how compute, networking, storage, orchestration, and inference fit together. Basic familiarity with IT, cloud, networking, or data center concepts will be helpful, but no previous experience with NVIDIA GPUs is required. You do not need a programming or data science background to follow the book; the focus is on infrastructure concepts, technology boundaries, operational decisions, and platform trade-offs.