New preprint: CUDAHercules, a benchmark for hardware-aware, expert-level CUDA optimization with LLMs.