Community Benchmark Board

August 12, 2026 ยท View on GitHub

This page tracks benchmark reports shared by users. Submit results with the Benchmark result issue template.

For trading-workflow validation rather than tensor throughput, use public_data_validation.md. That path reports walk-forward baselines, transaction costs, slippage, turnover, and drawdown. It also writes a copy-ready submission.md and machine-readable JSON files for community reports.

How to Submit

python -m pip install -e '.[dev]'
make benchmark

Then open a benchmark issue and paste the printed Markdown table.

For larger panels:

python scripts/benchmark_tensor_factors.py \
  --device auto \
  --n-dates 1500 \
  --n-stocks 3000 \
  --window 20 \
  --repeat 5 \
  --warmup 2

Results

ContributorCommitOSPythonPyTorchCUDACPUGPUCommandNotes
Maintainer44c6777Windows 11 10.0.262003.14.42.11.0+cpuunavailableIntel Core i7-1255U, 10 cores / 12 threadsnonepython scripts/benchmark_tensor_factors.py --device auto --n-dates 750 --n-stocks 1000 --window 20 --repeat 5 --warmup 2CPU-only Windows report
Maintainerd3a99b6macOS 26.5.1 arm643.9.62.8.0unavailableApple M5, 10 cores, 32 GB RAMnonepython scripts/benchmark_tensor_factors.py --device autoCPU-only report on MacBook Air

Maintainer CPU Baseline: Intel Core i7-1255U

Environment:

  • Commit: 44c6777
  • Machine: Intel Core i7-1255U, 10 cores / 12 threads
  • OS: Windows 11 10.0.26200
  • Python: 3.14.4
  • PyTorch: 2.11.0+cpu
  • CUDA available: false
  • CUDA version: unavailable
  • Synthetic panel: 750 dates x 1000 stocks
  • Warmup / repeat: 2 / 5
DeviceCaseMeanStdPeak CUDA memory
cpucs_rank(close)23.4 ms2.5 ms-
cputs_mean(close,20)21.1 ms1.7 ms-
cputs_rank(close,20)71.3 ms6.3 ms-
cputs_corr(close,returns,20)151.2 ms10.2 ms-
cpuewma(close,0.05)50.4 ms15.6 ms-
cpucompute_legacy_set(6 factors)287.0 ms15.8 ms-

The Windows and macOS baselines use different Python and PyTorch versions, so they are reproducible machine snapshots rather than a controlled CPU ranking. Use the same command and software environment for strict hardware comparisons.

Maintainer CPU Baseline: Apple M5 MacBook Air

Environment:

  • Commit: d3a99b6
  • Machine: MacBook Air, Apple M5, 10 cores, 32 GB RAM
  • OS: macOS 26.5.1 arm64
  • Python: 3.9.6
  • PyTorch: 2.8.0
  • CUDA available: false
  • CUDA version: unavailable
  • Synthetic panel: 750 dates x 1000 stocks
  • Warmup / repeat: 2 / 5
DeviceCaseMeanStdPeak CUDA memory
cpucs_rank(close)7.4 ms151.8 us-
cputs_mean(close,20)3.6 ms356.1 us-
cputs_rank(close,20)11.7 ms504.5 us-
cputs_corr(close,returns,20)20.8 ms618.6 us-
cpuewma(close,0.05)3.2 ms87.9 us-
cpucompute_legacy_set(6 factors)60.0 ms807.6 us-

What Makes a Good Benchmark Report

  • The command is copy-pasted exactly.
  • The commit SHA is included.
  • CPU and GPU names are included.
  • CUDA availability is stated.
  • The result table is pasted without editing numbers.
  • Any unusual conditions are documented, such as shared GPU, thermal throttling, or low memory.

Why This Helps

Benchmark reports help users answer practical questions:

  • Does GPU help for my panel size?
  • Which operations dominate runtime?
  • Is performance sensitive to PyTorch or CUDA version?
  • What hardware is enough for a student or research workflow?