The public release retains the core single-card concurrency/capacity milestones, their resource traces, the relevant ROCm stress reference and the completed ROCm 10.1 repeat cells. Other exploratory results are summarised here; their per-turn CSVs are omitted. Restored exploratory SVGs remain in the secondary graph archive, with per-graph exclusion reasons. Private source evidence is preserved.
Compaction-depth sweeps, storage/loop/tier experiments, scheduling and speculation variants, instrumentation checks, tainted/failed attempts and early agent-harness probes addressed tuning or boundary questions. They do not provide a matched ROCm comparison or the central 131k single-card concurrency result.
The historical one- and two-agent references are short growth tests; the RAM16 three- and four-agent milestones use different histories. The retained older/newer four-agent stress pair is also unmatched. None is presented as a controlled scaling curve.
Machine-readable exclusion summary
| Coverage | Excluded entries | Why omitted |
|---|---|---|
| Telemetry slots | 30 | Exploratory tuning, instrumentation, storage, compaction, capacity boundaries or separate harness trials; includes one empty slot. |
| Historical benchmark cells | 42 | Redundant or unmatched workload/configuration variants outside the focused comparison. |
Home · Overview · ROCm 10.1 results · Charts