chapter twelve

12 Now what?

 

This chapter covers

  • Topics we deliberately skipped and when they matter
  • When to stay above CUDA and when to drop below frameworks
  • Serving systems and why end-to-end performance matters
  • Compiler and generated-kernel workflows
  • Hardware targets beyond datacenter NVIDIA GPUs
  • How to keep learning without chasing every new tool
  • Resources for keeping pace with the ecosystem

By this point, you’ve gone from writing your first cudaMalloc to implementing tensor core operations, fused attention kernels, quantization strategies, and multi-GPU distributed systems. You can now read CUDA code in the wild, understand the kernel-level decisions, and write your own kernels when needed.

I kept this book deliberately narrow in scope. We focused on CUDA, the fundamentals, and the building blocks that do not change every few months. I did not try to cover every framework, library, serving stack, compiler, or hardware target in the AI ecosystem because that would have obscured the core ideas.

This final chapter is about what comes next. Some of that is technical: which tools to learn, which abstractions to trust, and where custom CUDA still matters. Some of it is professional: how to stay useful as compilers, generated kernels, and production inference systems absorb more of the work that used to require hand-written kernels. Think of this as a roadmap for continued study, not a checklist to finish immediately.

12.1 What This Book Intentionally Left Out

12.1.1 Datacenter-Level Considerations

12.1.2 Convolution-Heavy Architectures

12.1.3 The Blackwell TCGen05 Instructions

12.2 Choosing the Right Abstraction Layer

12.2.1 Triton: Python-Level GPU Programming

12.2.2 torch.compile

12.3 Serving Systems and End-to-End Performance

12.3.1 vLLM

12.3.2 SGLang

12.3.3 TensorRT-LLM

12.3.4 ONNX

12.4 Compiler and Generated-Kernel Workflows

12.5 High-Level Frameworks and Languages

12.5.1 JAX

12.5.2 Mojo

12.6 CUTLASS and Template Libraries

12.6.1 Python CuTe DSL

12.6.2 CUTLASS Profiler and Auto-Tuning