Story perspectives

Tawa Boosts GPU Performance with Automated Code Optimization

10/18/2025

Summary

Story summary

Tawa, an automated compiler, optimizes GPU performance by generating warp-specialized code from high-level tile-based programs.
It uses asynchronous references to simplify warp-level communication and improve data flow.
Experiments on NVIDIA H100 GPUs show Tawa delivers up to 1.2x speedups over Triton and matches hand-optimized kernels, reducing programming effort.
Advancements such as cooperative compute warp groups and persistent kernels further boost performance without manual restructuring.

1 of 1

Story summary

Tawa, an automated compiler, optimizes GPU performance by generating warp-specialized code from high-level tile-based programs.
It uses asynchronous references to simplify warp-level communication and improve data flow.
Experiments on NVIDIA H100 GPUs show Tawa delivers up to 1.2x speedups over Triton and matches hand-optimized kernels, reducing programming effort.
Advancements such as cooperative compute warp groups and persistent kernels further boost performance without manual restructuring.