This folder contains the selected results only. Version assignment follows the reviewed catalogue; older traces use profile/date evidence, not per-run immutable image IDs.
What was excluded and why. Different workload/configuration histories are not a controlled ROCm A/B.
Four completed GPU A single-card cases · Configuration · Side-by-side requirements and conclusions
Eight completed follow-up cases: single vs dual · Review limits
The seven-cell dual export uses an 8 GiB tier. The eight-cell GPU A/dual follow-up uses 10 GiB and is a separate campaign. The report contains serving settings, preemption counts and first coding-fleet outcomes.
The follow-up CSV counts timed turns: GPU A four-agent = 114; dual two-agent = 63. Including one empty completion in each gives 115/64 completed responses in the acceptance ledger. TTFT excludes those empty completions. Earlier review-notes JSON records the scope of the original export; the later report adds handover evidence without changing the exported data.
The eleven ROCm 10.1 games are published separately; raw session transcripts remain private. The raised-budget single-card retry completed with four clean deliveries in 75:23.
Per-agent fleet outcomes records all twelve agent attempts, including the missing first-attempt game. It is transcribed from the completed local campaign report, not a new session parse. Blank usage/timing fields were not available precisely in that report; they are not zero. Eleven games are linked by fleet and agent.
Home · Overview · ROCm 10.1 results · Charts