1\section{Conclusion}2\label{sec:concl}34Can high-frequency econometric workloads be moved to an FP32-only GPU without losing accuracy, inference or reproducibility? For the estimators studied here the answer is yes on accuracy and reproducibility and ``sometimes'' on speed. Compensated double-float accumulation, chunked FP64 combination and mixed-precision design keep Metal results within $10^{-6}$ of FP64 for realized measures and regression statistics and within $2\times10^{-5}$ for batched GARCH likelihoods, bitwise reproducible across runs and thread counts, with Monte Carlo coverage and reference comparisons indistinguishable from the CPU path. The speed answer is specific: the fused realized kernel beats an 18-thread CPU beyond about $4\times10^4$ elements (2--3.4$\times$ on raw reductions, 1.1--1.4$\times$ once the estimator's own FP64 post-processing is counted), batched GARCH likelihood grids gain 5--8$\times$, rolling sums and single-vector recursions lose, small panels tie, and the pipeline from Parquet to estimates is governed by column projection and calendar logic rather than by any kernel. Across eleven Apple Silicon generations the GPU helps most where the CPU is weakest.56The practical lesson for econometric software on unified-memory machines is that the device is cheap and the conversion is not: a library that wants GPU speedups must keep FP32 copies resident across calls, fuse its passes, and reserve the GPU for problems whose parallelism is in the number of series or parameter vectors, not in the length of a recursion. The methodological lesson is that none of this requires new estimators---only explicit conventions, references and tolerances, which is where most implementations differ.7