appendix-a

appendix A  Installation and setup

 

This appendix provides everything you need to set up your environment and run the code examples from this book. Whether you use a local GPU or a cloud instance, these steps will get you up and running.

A.1 Prerequisites

Before installing anything, confirm your machine meets the hardware bar and that your OS and toolchain are supported.

A.1.1 Hardware requirements

Minimum requirements (for basic chapters):

  • NVIDIA GPU with CUDA support (compute capability 7.0 or later)
  • 8 GB or more of GPU memory (VRAM)
  • Examples: RTX 3060, RTX 3070, or newer

Recommended for most chapters:

  • NVIDIA GPU with tensor cores (compute capability 8.0 or later)
  • 16 GB or more of GPU memory
  • Examples: RTX 3090, RTX 4090, A100, H100

Chapter-specific requirements:

  • Chapters 1 to 6—GPU with compute capability 7.0 or later
  • Chapter 7 (WMMA tensor cores)—Volta or later (V100, RTX 20+ series)
  • Chapter 7 (WGMMA tensor cores)—Hopper architecture (H100)
  • Chapter 10 (distributed, single node)—8× GPU server (H100 recommended)
  • Chapter 11 (CUTLASS)—Hopper (H100) for most examples; an SM100-based B200 or GB200 GPU for the FP4 example, which targets sm_100a and TCGen05 block-scaled tensor core instructions

A.1.2 Software prerequisites

You will need the following on the host before installing CUDA:

A.2 Core installation

A.2.1 Step 1: CUDA Toolkit

A.2.2 Step 2: Python environment

A.2.3 Step 3: Additional libraries

A.3 Cloud GPU options