> numpy can't really do like native GPU execution

I'd be interested to see where GPU code beats NUMPY's SIMD implementation, which is really

> features that numpy can't really do like native GPU execution

we use numpy + jax for that; works well

I wonder if the Python-to-Mojo overhead is significant enough where language-native bindings make more sense. I'd be interested in seeing some benchmarks.