Skip to content

MPSBoost 0.5.0 Performance Report

This report summarizes performance evidence for the 0.x hardening line. It is deliberately conservative: checked-in S4/S6 numbers are historical pre-v2/v3 optimization baselines, while 0.5.0 focuses on executable customer workflows, transparent backend selection, and documented performance boundaries.

Evidence Available

  • S4 histogram benchmark on Apple M2 Ultra, Python 3.13.5, MPSBoost 0.2.0a0.
  • S6 end-to-end regressor benchmark on Apple M2 Ultra, Python 3.13.5, before later v2/v3 cleanup.
  • Real-world acceptance reports for built-in, cached, and opt-in datasets.
  • CI package tests on hosted CPU runners and self-hosted real Metal GPU runners.

Historical MPS Boundary

The checked-in S6 report shows the expected shape:

Scenario Rows Features CPU median (s) MPS median (s) Speedup
gbdt-medium 16,384 32 0.067630 0.089597 0.755x
gbdt-wide 16,384 128 0.283176 0.249502 1.135x
gbdt-large-wide 32,768 256 1.031002 0.633006 1.629x

These are not final 0.5.0 speed claims. They are retained to show the workload boundary: small or synchronization-heavy work can be slower on GPU, while wider/larger histogram workloads can benefit from MPS.

Current 0.5.0 Policy

  • Use device="auto" for ordinary users.
  • CPU is the correctness oracle and the preferred path for small data, branch-heavy anomaly detection, and latency-sensitive ranking.
  • MPS is an acceleration backend for workloads large enough to amortize launch, transfer, and synchronization overhead.
  • Explicit external portable policies record requested and effective backends. If the external runtime is not active for the estimator, MPSBoost warns and continues through native CPU compatibility rather than stopping the workflow.
  • Linux CPU/CUDA performance depends on the selected external sklearn/XGBoost/CUDA stack and is reported as that external backend, not native MPSBoost CPU/MPS.

Before 1.0.0

The final stable release needs a new full performance audit on current HEAD:

  • train time and predict time across built-in, cached, and opt-in real-world datasets;
  • peak memory;
  • model size;
  • wheel size;
  • CPU/MPS applicability boundary;
  • external backend attribution for Linux CPU/CUDA paths;
  • raw command lines, machine details, package versions, and artifact hashes.