A graphics programmer's roadmap: master the GPU, the CPU, and write your own path tracer

Modern rendering splits into two hard-to-learn-simultaneously halves: the CPU side (explicit APIs like DirectX12, Vulkan, or Metal, plus the engine plumbing to load assets and drive the pipeline) and the GPU side (the math of lighting and shading, techniques like shadows and ambient occlusion, and a feel for what runs fast in real time). The advice is to focus on one and lean on something simpler for the other — use OpenGL, WebGL, or an existing engine if you want to concentrate on shading, or just get a triangle and then a mesh on screen if you want to concentrate on the API.

The recommended proof-of-competence portfolio is concrete: an engine-like real-time renderer that loads models and textures, applies physically based rendering (PBR) with a couple of effects and a movable camera in C++, plus a standalone path tracer that outputs a photorealistic image. The strongest version folds the path tracer into the same renderer as a reference mode, using it to validate the real-time PBR output — and earns extra credit for explaining where the two diverge and how to close the gap while staying real time. Suggested learning resources include “Ray Tracing in One Weekend,” LearnOpenGL’s PBR theory pages, Google’s Filament docs, and the PBRT book, all free online. The math needed is modest (linear algebra, basic trig, a little calculus); C++ and HLSL remain the default languages, with Rust and WebGPU present but not yet mainstream in job listings.

The author, a working graphics programmer, frames all of this against skepticism about the current AI hype cycle, predicting the pendulum will swing away from ML and that hype-chasers will move on to quantum computing next. He still finds narrow uses for LLMs — discussing math and papers, spotting possible bugs — but not for writing code he’d then have to fully understand anyway. ML’s genuine value, he argues, lies in its fitting and optimization techniques, and today’s LLM era is best understood as a dress rehearsal for whatever real machine intelligence eventually arrives.