Performance Engineer Take-Home — Kernel Optimization

Performance Engineer track replaces the standard coding phone screen with a 2-hour AI-allowed take-home: optimize a mocked GPU-kernel-style core function. Scoring is by measured speedup; passing scores reported in the 600/1000 range with ~8× speedup. Requirements 2-hour take-home. AI tools (Claude, ChatGPT, Copilot) are explicitly allowed — this is the rare Anthropic round that encourages them. The target is a mocked kernel function that resembles a real GPU kernel: nested loops over tiles, arithmetic on a buffer, output reduction. Optimizations expected: loop unrolling, memory coalescing / better stride access, operator fusion, vectorization, eliminating redundant work. Scoring is based on a benchmark harness (speedup × baseline = score). Reported pass bar: roughly 600/1000 corresponding to ~8× speedup over the baseline. Notes The take-home is the gate into the Performance Engineer onsite — clearing it skips both the standard recruiter phone-screen technical chat and any live coding screen. Don't optimize blindly: read the kernel structure, identify the bottleneck (memory access vs. arithmetic), then iterate. The score is monotonic in measured speedup, so iteration time matters more than elegance. AI tools are useful for generating candidate rewrites; the candidate's value-add is recognizing which rewrites are actually faster on the harness. Preparation Read the standard GPU-kernel optimization references (Nvidia's CUDA C++ Best Practices, PMPP chapter on coalescing). Memorize the ladder: coalescing → tiling → unrolling → fusion → occupancy. Practice profiling a CPU baseline with perf or cProfile so you can identify the hot inner loop quickly. Have a saved scaffolding for benchmarking + correctness-checking so you don't burn time on infrastructure inside the 2-hour window.