The Cost of Irregularity: CUDA C++, Rust, and Triton
A comparison of CUDA C++, Rust, and Triton in GPU programming. Performance differences in irregular workloads are analyzed.
This study compares GPU languages by implementing a hash-blocked TSDF fusion kernel in CUDA C++, Rust via NVIDIA's cuda-oxide, and Triton. The performance results indicate that while all three languages perform similarly in regular workloads, Rust closely matches hand-written CUDA C++ in irregular workloads, whereas Triton is significantly slower. These discrepancies arise from specific limitations in language expression rather than mere performance ratios.