From Naive CUDA to Triton: A Systematic Evaluation of AI-Era HPC Operator Development

High-performance operator development is central to AI-driven scientific computing, yet manual CUDA optimization remains time-consuming and expertise-intensive. Large language models (LLMs) and domain-specific languages such as OpenAI Triton have recently emerged as promising alternatives for GPU kernel development. This paper presents a systematic evaluation of these programming paradigms using a four operators considering computational intensity and memory access regularity. We study four representative kernels from real HPC applications: Stencil (miniGhost), SpMV (Siesta), GEMM (GAMESS), and N-Body molecular dynamics.

Experiments on an NVIDIA RTX 3090 show that Triton performs strongly on regular compute-bound workloads, outperforming cuBLAS by 1.28 × for GEMM through auto-tuning and software pipelining. However, its performance degrades on irregular access patterns and collapses in all-to-all interaction patterns: Triton-based N-Body is 13.5 × slower than naive CUDA. We further show that high SM occupancy does not necessarily imply high performance, and that compute-unit utilization provides a more reliable indicator of kernel efficiency. These results characterize where LLM-assisted CUDA and Triton are effective, where expert-level optimized CUDA remains necessary.