Meta needed to bring new model operations to its own accelerator without writing each one in low-level C++. Triton helped the team do that. An experiment with a 3D map shows where the choice changes. On the same GPU, Rust and CUDA C++ find space for new observations much faster.
