From CUDA to MLX: How K-Search Brings Decades of Kernel Expertise to Apple Silicon

Researchers from UC Berkeley Sky Lab built on K-Search, an evolutionary kernel search framework that uses AI to optimize GPU kernels, and extended it with a backend for Apple's MLX framework. They developed a novel structured CUDA-to-MLX translation layer that converts existing CUDA kernels into high-quality kernels for Apple Silicon. The approach reached 0.97x speedup compared to native MLX Attention kernel and up to 20x prefill speedup over community mlx-lm implementation on the Mamba SSM kernel. K-Search uses an iterative optimization loop where an LLM reasons about optimizations, generates candidate kernels, compiles and benchmarks them on real hardware. The translation layer includes concept mapping tables, MLX-specific hints, and reusable assertions that preserve expert kernel behaviors. The work demonstrates that AI-driven kernel search can transfer optimization knowledge across hardware ecosystems without starting from scratch.