This gallery includes selected results only; exploratory graphs are available separately in the unchosen archive; their per-turn CSVs remain omitted. See what was excluded and why. Charts use descriptive experiment names and separate ROCm folders. Version assignment follows recorded build/runtime evidence; unknown versions stay separate. Different workloads and settings are not a controlled ROCm comparison. Prompt-credit traces are diagnostic, not validated prefill throughput.
ROCm base-image four-agent stress run</summary>
The 17 October journal explicitly identifies this stress workload as ROCm 10.1.
Sustained four-agent load on the newer runtime stack. Compare with older-stack data only as an indicative comparison because configuration and depth caps differ; this is a stress qualification.
</details>
Exploratory cases within this recorded software profile; workloads and settings vary. These bars are not a controlled comparison.
Play all eleven games · Per-agent outcomes CSV. Both final fleets delivered four games; budgets and client settings differed. The failed first single-card attempt remains in the report.
Short-output decode estimates are not fleet throughput or real-agent productivity. Source data.
Different tier budgets, utilisation and image patches prevent an isolated ROCm-version conclusion.
See the seven completed dual-card text cases. Eight GPU A/dual follow-up cells have completed; see the latest results and README requirements and results and recorded configuration. The campaign report records pressure telemetry and both first fleets; the single-card retry completed with all four agents delivering and exiting cleanly in 75:23.
Single-card RAM16 tier, four agents at 131k target</summary>
ROCm 7.14 by recorded active profile and 6–8 October collection date; no immutable image ID captured per trace.
Qualification run for the 16 GiB shared-memory tier with four agents on one card. It is the key smooth-operation milestone; read error count, turn completion, TTFT by depth and memory pressure together.
</details>
Single-card RAM16 tier, three agents at 150k target</summary>
ROCm 7.14 by recorded active profile and 6–8 October collection date; no immutable image ID captured per trace.
Follow-up capacity point with fewer agents and deeper context. Compare sustained completion and tail behaviour with the four-agent milestone.
</details>
Host-memory trace for the RAM16 four-agent milestone</summary>
ROCm 7.14 by recorded active profile and 6–8 October collection date; no immutable image ID captured per trace.
Companion resource trace for the central four-agent single-card result. Check swap stays controlled while the shared-memory tier is active.
</details>
Host-memory trace for the RAM16 three-agent follow-up</summary>
ROCm 7.14 by recorded active profile and 6–8 October collection date; no immutable image ID captured per trace.
Resource companion for the deeper three-agent test. Compare memory headroom and swap with the four-agent tier run.
</details>
Exploratory cases within this recorded software profile; workloads and settings vary. These bars are not a controlled comparison.
The short one/two-agent references and the three/four-agent RAM16 milestones have different histories/settings.