I have a stress test written in C: 100,000 cubes flying around in space. The CPU rebuilds every cube’s 4x4 transform matrix on every frame and sends all of them to the GPU. That is around 900,000 sine evaluations and 6 MB of matrix data per frame, and after that the GPU still has to draw 3.6 million triangles. So the frame is half CPU work, half GPU work. I ported it to Clojure and wanted to see how close I could get to the C version’s FPS. I should say up front that I did not do the optimization work alone: I paired with Claude Code on it, and most of the digging in this post (the benchmarks, the JIT logs, the failed attempts) comes from that session.