A case for learning GPU programming with a compute-first mindset – Maister's Gr (themaister.net)

This piece argues that newcomers should learn GPU programming with a compute-first, debugger/profiler-first mindset rather than starting with traditional raster graphics “Hello Triangle” exercises. The author says most modern GPU work — especially relevant to AI/ML — is compute-centric, so teaching compute shaders, tooling, and the GPU execution model first gives faster, more practical returns. They recommend a mid-level Vulkan-focused approach (the author’s Granite library as an example) to avoid raw-API overwhelm while keeping close to real Vulkan behavior, and to delay learning the large fixed-function raster API until fundamentals are understood. Key technical practices: use Vulkan compute (Vulkan 1.2 for buffer device addresses) and Vulkan GLSL as the shading language; adopt a debugger-centric workflow with RenderDoc + SPIRV-Cross to decompile, edit, and replace shaders (including debugPrintfEXT) for rapid iteration and aggregate debugging; inspect machine ISA (RADV/Mesa recommended) and study the dispatch/workgroup/invocation hierarchy to internalize how thousands of threads map to SIMT hardware. The author also suggests deferring explicit synchronization and starting with single-dispatch experiments to simplify learning. For AI/ML engineers, this path accelerates understanding of kernel behavior, memory/addressing, and performance pitfalls that directly impact model training/inference optimizations on GPUs.