AI agent ports and optimizes CUDA attention kernel to CuTeDSL in an afternoon, 1.2x faster

knowrohit07 · x · 2026-09-17

In a new article, maharshii reports his AI agent ported and optimized a CUDA C++ attention kernel to CuTeDSL in a single afternoon — the result ran 1.2x faster than the original kernel and 2.8x faster than cuDNN.

Commenter knowrohit07 adds that getting comfortable with tensor dialects (lower-level than DSLs or C++) pays off: he has written kernels directly in MLIR and ported parts of the CuTe DSL, arguing DSLs generally have poor debugging experiences.

Original post →

More from coding & agent

coding & agent channel →