Real GPU experiment results

These reports are produced by actual inference runs. The experiment recipes include historical RTX 3080 reproduction and a separate H100 suite with user-selected models and GPUs.

• RTX 3080 local results
• H100 results archive — reserved for real H100 executions; currently no H100 measurements are published.

The reports open with a concise throughput/duration summary, visible confidence, and best-observed settings. All trials remain visible with their average benchmark execution duration; expandable details preserve all recorded trial data. Workload comparisons, validation evidence, charts, and methodology are collapsed initially. Missing or incomparable measurements remain unavailable.

Each result includes its exact experiment, score/measurement exports, and an evidence bundle. Compare baseline and candidates within the same model and workloads. The RTX smoke workloads demonstrate reporting; the H100 large-model scenario measures gains separately and requires fresh candidate validation. Different models do not establish cross-GPU speedup claims.