Skip to content

H100 results archive

No H100 experiments have been executed or published yet. This folder is reserved for real results from the rented H100 GPU machines, not simulated or RTX 3080 measurements.

Use the H100 scenario suite with your chosen models and GPU IDs: parallel independent trials, one server spanning 2+ GPUs, and large-model tuning followed by separate validation. H100 models do not need to match the historical RTX 3080 model.

Store each actual run under <date>/<experiment>/<run-id>/, with:

  • report.html, result.json, results.csv, and experiment.yaml;
  • model/YAML checksums and software/hardware provenance;
  • both search and validation reports for large-model tuning, with the baseline rationale and per-workload improvements or regressions;
  • normalized results, manifests, raw benchmark output, and diagnostic logs in an evidence archive, with public redactions disclosed.

Add links here after execution, including unsuccessful runs. Do not copy the local report into this folder or label unrun configurations H100-validated.