UC Berkeley's K-search: Auto-Translating CUDA Kernel Optimizations to Apple's MLX

berkeley_ai · x · 2026-07-30

As the AI hardware ecosystem fragments, transferring mature kernel optimizations (like NVIDIA's CUDA) to new chips (like Apple Silicon) typically requires extensive manual rewrites.

To solve this, UC Berkeley and IBM Research introduced K-search, an evolutionary kernel search framework. It automatically translates decades of CUDA optimization knowledge into native strategies for Apple's MLX framework. Benchmarks show their translated attention code achieves 0.97x the speed of Apple's native implementation. The project is open-source.

Original post →

More from Infra

Infra channel →