Get the latest tech news

Show HN: Phobos – A tiny scale-free kernel language with tile-DAG support

I built a tiny GPU kernel language from scratch to learn the stack: hand-written PTX, then MLIR down to NVPTX, reaching 76% of cuBLAS SGEMM on an old RTX 2080 SUPER. First-class tiles then let the same kernel compile to a multi-node tile-DAG (a design I’ve only run on one machine).

None

Get the Android app