Research Projects
Selected public research and software infrastructure. Projects still under anonymous review are not detailed here.
A benchmark of 114 HPC workloads across 16 computational motifs evaluating LLM-driven optimization in serial CPU, OpenMP and CUDA. It combines reproducible builds, correctness checks and the Fast@k / Speedup@k metrics. The study finds that reliable optimization becomes harder as required speedup increases, particularly on CUDA.
A controlled benchmark comparing 45 operators in Triton and cuTile on NVIDIA B200. Matched semantics, autotuning, NCU/PTX diagnosis and LLM-generated kernels expose workload-dependent compiler and hardware performance differences.
An ongoing framework for performance prediction and optimization in HPC applications. The open benchmark infrastructure spans GPU microbenchmarks, proxy/mini-applications and production scientific codes, with standardized build, execution, timing and validation workflows. Performance prediction and optimization components remain active research goals.
An earlier study of 24 HPC programs across nine computational domains, comparing LLM-driven serial and parallel optimization against compiler-based tools. Its correctness and performance findings motivated the expanded HPC-Bench evaluation.