Valente Benchmarks: A Universal Registry for Normalized Computational Complexity

Valente Benchmarks (VB) reimagines traditional computational complexity benchmarks as a dynamic, predictive infrastructure. By normalizing runtimes to a standardized Computational Unit (CU) based on sustained double-precision floating-point operations per second (GFLOPS), and employing Gaussian Process (GP) regression for uncertainty-aware interpolation, VB enables cross-platform, cross-era comparisons and conservative forecasting. Initial empirical validation on six diverse hardware architectures for 3-SAT instances (n=50) reveals statistical invariance in scaled times (one-way ANOVA: F(5,30)=1.23, p=0.31), with a mean T_scaled of 158.8 ± 0.48 s/CU. Acknowledging the limitations of standard kernels for extrapolating to large problem sizes—particularly near phase transitions—VB emphasizes interpolation within validated data regimes while archiving raw measurements to fuel advanced kernel development. This open registry invites contributions from machine learning, algorithms, and complexity theorists to build a shared, reproducible resource for benchmarking NP-hard problems and beyond. 10

Keywords: computational complexity, Gaussian processes, benchmarking, reproducibility, SAT solvers, phase transitions.