What Is the GPU Speedup Measured Against? Library-Version Lock-In and Core-Count Confounds in CPU-versus-GPU Dataframe Benchmarks

A GPU speedup is a ratio, so what it means depends on its CPU denominator. We re-examine the CPU baseline of a pre-registered benchmark of Polars and Apache Spark, each on CPU and GPU, run as four separate campaigns on single-GPU A100, H100, GH200 and B200 hosts with a TPC-H-derived workload (not a TPC-H result). Version lock-in. The Polars GPU engine requires Polars below 1.45, so its pre-registered speedup is measured against Polars 1.44.2 on the CPU. In the pre-registered bridge, Polars 2.0.0 took 0.62–1.03 of 1.44.2’s time on the baseline suite and 0.46–0.52 on the complex suite. Every 95% interval lay below 1 on every host except B200 baseline SF1, where it includes 1. Measured against 2.0.0 instead (exploratory), the H100 speedup at SF300 falls from 2.81× (95% CI 2.64–2.99) to 2.09× (95% CI 2.05–2.14), or to 2.05× (95% CI 1.92–2.18) after correcting for a measured session bias. On the GH200 at baseline SF30, the GPU engine is faster than Polars 1.44.2 but slower than Polars 2.0.0. Equal threads, unequal cores. The same 26 threads occupy 13 physical cores on the H100 and B200 (two hardware threads per core) but 26 on the A100 and GH200. The GH200’s Polars GPU speedup at SF300 was 1.78×, against 2.81× on the H100, mainly because its CPU side is faster (exploratory factorization). The CPU model and architecture also differ, so core count is not isolated. We end with a reporting checklist for CPU-versus-GPU benchmarks. Workload notice. The workload is derived from the TPC-H Benchmark and as such is not comparable to published TPC-H Benchmark results, as the results of this workload do not comply with the TPC-H Benchmark Specification.