This research compares ROCm 7.14 and ROCm 10.1 using the vllm-radiance repository on Radeon AI PRO R9700 hardware, serving the Swift-1.5 Qwen3.8 27B INT4 checkpoint for coding agents.
A central workload asks how many concurrent agents can each sustain roughly 131k tokens of context on one R9700 plus host RAM, with useful latency and successful completion. Single-card and dual-card configurations are reported separately. The ROCm 10.1 eight-cell matrix and all three fleet attempts are complete. Dual delivered four games in 30:10; the revised-budget single-card retry delivered four in 75:23. Configuration differences prevent an isolated ROCm-version conclusion.
The measured ROCm 10.1 single-card 10 GiB-tier profile supported 1–2 deep agents with useful latency, then saturated at 3–4. The dual deployment completed the four-agent matrix and coding fleet with substantially better latency and completion.
The recorded campaigns use different settings. Read the configuration and provenance before comparing results; a configured 131k ceiling alone does not establish sustained simultaneous operation at that depth.
Play the eleven ROCm 10.1 games, grouped by fleet and attempt, or browse the historical collection. Explore the Falling Blocks game gallery, the real-agent task record and the artefact provenance.