Berkeley AI Research uses K-Search to translate CUDA kernel expertise to Apple Silicon's MLX

Researchers built a novel structured CUDA-to-MLX translation layer on the K-Search framework to adapt existing CUDA kernels for Apple Silicon's MLX framework. This approach achieves near-expert level performance, showing a 0.97x speedup over the native MLX Attention kernel and up to a 20x prefill speedup over the community mlx-lm implementation on the Mamba SSM kernel. This is significant because MLX, adopted widely since late 2023 for local AI inference on Apple Silicon, often lacks performance-critical kernels that the NVIDIA ecosystem relies on.