[lvxianchao]

Problem and result

docs/AMD_HIP.md marks the rx6000 Linux package "built, not tested on a real RX 6000 card", and asks for a run on such a card with the end of the engine log. That run was missing; this PR adds it.

Before: no report existed for a real RX 6000 card with the released package, and the RDNA2 section's numbers came only from source-built engines.

After: the released rx6000 package runs on a real Radeon RX 6900 XT (gfx1030) — the standard protocol passes 12 of 12 sanity checks (including all four tool-call checks) at prompt 908.5 / 1,128.9 tok/s and decode 39.7 / 40.2 tok/s (4K / 16K), with IQ3_S at 262,144 tokens of context and pictures on.

Related discussion

None open. CONTRIBUTING.md allows benchmark reports to go straight to a PR.

Approach

Results only, as CONTRIBUTING.md asks for. No engine code, default or behaviour is touched.

• bench/results/2026-10-30-standard-navi-21-rx-6800-6800-xt-6900-x-0f8809de/ holds what strata bench run wrote: README.md, results.json, outputs.jsonl, config.json, provenance/, plus strata-iq3_s-session.log (the engine log tail that docs/AMD_HIP.md asks for).
• community-notes.md in the same folder holds the measurements made outside the protocol, each labelled as such and not comparable with standard results: the STRATA_HIP_PROMPT_F16=1 before/after, the flat cost of context size from 64K to 512K, prompts up to 300K tokens, pictures on gfx1030, and three negative results so nobody repeats them.
• One bullet is added to the list in docs/COMMUNITY_BENCHMARKS.md.

Validation

Passed: strata bench run --url http://127.0.0.1:8080 --track custom --track-notes "…" --quantization IQ3_S, engine 0.1.54 linux-x86_64-hip-rx6000 (release package, no source build), run 20261030T154543Z-0f8809de. 12 of 12 sanity checks (instruction 4/4, retrieval 4/4, tool calls 4/4), no failed request, 3 min 54 s of measured work after the model loaded. Everything is in the folder above.

Failed: none observed.

Not run: the Windows rdna2 package; any quantization other than IQ3_S (IQ3_XXS was not downloaded); bare metal; the concurrency and soak suites; --layer-split (single card); per-request clocks, power and temperatures.

Limits

• The machine is a QEMU/KVM guest on Proxmox VE 9.2 with the GPU passed through (vfio-pci, x-vga=1, q35 + OVMF, cpu: host); the model files sit on the host's NVMe behind virtio. The absolute numbers are therefore not bare-metal numbers. PCIe is Gen4 x16 (LnkSta: Speed 16GT/s, Width x16, Resizable BAR 16 GB).
• 6 of the 9 performance requests were truncated at the 512-token cap because reasoning used the budget. That is the cap, not a failure, and results.json records it.
• The system ROCm is 6.3.0 and is not used by the package (it carries its own HIP runtime). Worth stating because the RX 6000 docs point at a system ROCm 7; this run shows a released package needs no system ROCm at all.
• The rows in community-notes.md are one request per cell, not medians, except the ten STRATA_HIP_PROMPT_F16 on-arms. They answer the questions this report raises; they are not a second benchmark.
• I did not touch the "built and not tested on a real RX 6000 card" wording in docs/AMD_HIP.md; say the word and I will update it here or in a follow-up, whichever you prefer.

[github-actions]

Thanks for the pull request. To keep the tracker readable, every item needs an area prefix in the title and the parts of the pull request template filled in. This one is missing:

• Fill in "Problem and result": what triggered the change, and how it behaves before and after.

Edit the pull request (title and description) and this note goes away on its own. If it is still missing parts in 5 days it will be closed, which is not a rejection: fix it and comment /reopen.

Details: CONTRIBUTING.md

[github-actions]

Thanks, this follows the format now.

[Niko1221]

Thank you for the benchmark. We have collected these results for the benchmarks page on the Strata website, so we are closing this pull request instead of adding the files to the repository. If you measure again, for example on 0.2.11, you are welcome to open a new report.

The Strata project