You got it right: This project exposes our tile programming model, whereas cuda-oxide exposes our lower level CUDA-like programming model. Our tile programming model is higher-level: It compiles to what looks like the CUDA-like programming model.

If you're using burn and candle, and you're writing custom kernels, you can probably write most of your kernels in cutile-rs and let the Tile IR compiler optimize your kernel.

That said, if you're used to writing CUDA, then there is a bit of a learning curve. We have tutorials available that walk you through how it works here: https://nvlabs.github.io/cutile-rs/0.2.0/index.html

Familiarity with numpy helps substantially (it's supposed to have a numpy-like feel), but if you're coming from CUDA and want to leverage the safety features this project provides, then you should jump straight to the "useful mental models" page, which touches on how this compares to CUDA: https://nvlabs.github.io/cutile-rs/main/guide/useful-mental-...