[Maddy55555]

The title must start with its area: [Engine], [GUI], [TUI], [Benchmark], [Docs] or [Other], for example "[Engine] fix decode stop on long prompts" or "[GUI] keep the model list after a tab switch".

Problem and result

Pilot data for the draft standard protocol strata-qwen38-thinking-v1-draft.1 on an Intel Arc Pro B70 32 GB (SYCL AOT). A complete run: 21/21 required requests, sanity 12/12, decode 44.4 / 37.9 / 34.8 tok/s and prompt 1,081 / 1,118 tok/s (short / 4K / 16K medians). Adds the report folder only; no engine changes.

Related discussion

none (results-only submission following docs/BENCHMARK_PROTOCOL.md §Submit and review; re-submitted under the tracker-reset guidance from the closing comment of #2922, this time on 0.2.2 with the new title format and the standard protocol instead of the community guide).

Approach

bench/results/2026-10-23-standard-arc-pro-b70-f329cb7f/ contains README.md, results.json, sanitized config.json, outputs.jsonl, a sanitized 273-line engine-log-excerpt.txt and three small provenance files (chat template, corpus build, corpus manifest). No weights, no packs.

Command: strata bench run --profile bench-profile-iq3_s.json --no-start --track custom --quantization IQ3_S --model-revision ed59f92082b1e93c0e96d60a8b11aab089b52f09 --source-revision de4b8e4d43b917e7706784d8bb445c9af86a3540 --tokenizer-revision de4b8e4d43b917e7706784d8bb445c9af86a3540 ... against an engine that had been reloaded five minutes earlier (fresh engine, single slot, nothing else on the GPU). Model: ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF IQ3_S, native pack built with tools/iq_pack.py, repo expert profile, engine 0.1.47 from the v0.2.2 source (SYCL AOT bmg-g31).

Validation

Passed:

• strata bench run (full protocol, revision 011c9e6, Ubuntu 26.04 VM with Arc Pro B70 xe passthrough): status complete, status_reasons empty, 21/21 required requests, sanity 12/12 (instruction 4/4, retrieval 4/4, tool_call 4/4), required stage 759.9 s.
• Provenance cross-check: both GGUF shards' SHA-256 equal the Hugging Face LFS OIDs of the quant repo at revision ed59f920 (checked against the HF API on 2026-10-23).
• Artifact SHA-256: results.json 475020939a4ececc2be8b8b3e3703e82c55e26315ad6713d614f33ad3fc5c1fb, README.md ac6bf86dce716b1b567998b08f5e39a2a42e21e1abd2931067177a06ad34c85e, outputs.jsonl 20f3d0e0039df5236a3c8fa40b0984bf18e6e5a63bd45199faa5457bfd3a3e72, config.json 4bc65232ebad8709921da21aa663daf10496c4c0c6ca9c61dd800367f8e60b5f, engine-log-excerpt.txt f8505d446c1fca218e8e8cdd4fc3b29ba66c47d5f253068c1a50f4c3f6e61812.

Failed: none observed. (The one warm-up request and six performance requests ended at the protocol output cap and are counted as truncated in request_counts, not failures.)

Not run: the optional concurrency stage (planned 0); GPU memory_bytes, power_limit_w and the driver/runtime version probes — the xe passthrough VM guest exposes nothing readable there, recorded as missing entries in results.json with reasons. No clocks pinned.

Limits

• QEMU VM with GPU passthrough: the reported CPU label (EPYC-Milan) is the QEMU label; the physical CPU is a Ryzen 7 5700G. The sysfs PCIe link is a QEMU stub; the engine's own probe measured 6.5 GB/s host->device (pcie_frac 0.18), the physical link is Gen3 at x16 width.
• Round-1 maxima (TTFT 106–156 s) are expert-cache warm-up after the five-minute-old reload; round-2/3 medians are steady state and reproduce a prior run within ~1 %.
• A background agent shares the same engine via loopback (declared via --background in the run command; single slot, it was idle during the measured requests).
• The four missing values above stay unresolved by tooling, not by configuration; readers comparing rows should keep them in mind.

[Niko1221]

Thank you for measuring this and writing it up. We are building a Strata website with a page for community benchmarks, and results like these will be added there when it launches, so we are closing this pull request here; the numbers are not lost. If your numbers change with a new version, you are welcome to send them again.