The versioned data folders contain selected results. Legacy identifiers are metadata for traceability. Times are relative; memory is GiB. Counter-derived rates need a prior sample. Prompt-compute credit is diagnostic, not measured prefill throughput.
| Experiment | Version | CSV |
|---|---|---|
| Single-card RAM16 tier, four agents at 131k target | ROCm 7.14 | CSV |
| Single-card RAM16 tier, three agents at 150k target | ROCm 7.14 | CSV |
| ROCm base-image four-agent stress run | ROCm 10.1 | CSV |
| Host-memory trace for the RAM16 four-agent milestone | ROCm 7.14 | CSV |
| Host-memory trace for the RAM16 three-agent follow-up | ROCm 7.14 | CSV |
Excluded coverage is summary-only.
| Dataset | Scope |
|---|---|
| Earlier dual growth | Seven completed cells; TP=2, 8 GiB CPU tier |
| GPU A growth | Four completed single-card cells; TP=1, MTP=4, 10 GiB CPU tier |
| Single/dual follow-up | Eight cells, including those same four GPU A rows; do not add both exports as independent samples |
| Single-card configuration | Checkpoint and reviewed launch settings |
TTFT is in seconds; decode estimates are tokens/s for short responses, not aggregate fleet throughput. Counts in these CSVs are timed turns. The later acceptance report additionally counts one empty completion in GPU A n4 and one in dual n2. Reported case wall time and harness elapsed time use different boundaries. See accounting and review limits.
Per-agent outcomes contains twelve rows across three fleet attempts, with eleven game files. deliverables_complete means all requested files existed; clean_delivery additionally requires exit 0. fleet_wall_seconds repeats the fleet wall time on each agent row and must not be summed. Output tokens are cumulative across requests; max_output_tokens_per_request is a per-request limit. Blank values are unavailable, not zero. Verification columns report target campaign checks; they do not claim fresh browser testing of these copies.
Home · Overview · ROCm 10.1 results · Charts