Workshop: International Workshop on RISC-V for HPC (RISCVHPC)
Authors: Peijun Li (University of Minnesota), Pei-Hung Lin (Lawrence Livermore National Laboratory), Hiren Patel (University of Waterloo), Caiwen Ding and Yang Zhao (University of Minnesota), and Maya Gokhale (Lawrence Livermore National Laboratory)
Abstract: The RISC-V Vector extension (RVV) offers rich architectural flexibility for data-parallel HPC workloads, but its vast microarchitectural design space is expensive to explore by cycle-accurate simulation, so machine-learned surrogates have become the standard substitute. Two questions that decide whether such a surrogate is useful are rarely examined: where its features come from, and how its generalization is measured. We study both on a gem5 model of the Ara vector unit, sweeping vector and memory-system parameters across TSVC2 kernels and training regression models on static, dynamic, and interaction features. We find that hardware configuration alone predicts CPI poorly, at 52.6% error under leave-one-kernel-out validation, while program features recovered from the compiled binary reduce this to 25.4%. More consequentially, the validation protocol dominates the reported accuracy: on identical data, features, and models, a random split reports 0.28% error against 23.2% when whole kernels are held out.
Back to International Workshop on RISC-V for HPC (RISCVHPC) Archive Listing
Back to Full Workshop Archive Listing