[Maddy55555]

Complete standard-draft.1 run on an Intel Arc Pro B70 (SYCL/XPU), same protocol and model as PR #3799, with two setup changes: engine 0.1.48 (v0.2.3 + max_tokens hotfix) and the CPU-expert path enabled (STRATA_VERIFY_NO_HOST=0 + STRATA_VERIFY_DEVICE_PLAN=0).

Protocol conformance

• Protocol: strata-qwen38-thinking-v1-draft.1 (draft)
• Run ID: 20261023T123010Z-a48c5a63, status complete, 0 failed requests, sanity 12/12
• Required requests 21/21 sent; warm-up excluded; no retries; single slot, loopback HTTP streaming

Machine

• GPU: Intel Arc Pro B70 32 GB (PCIe x16 Gen3 electrical — host CPU ceiling; engine probe 12.8 GB/s host->device -> pcie_frac 0.35, autotuned)
• Host: AMD Ryzen 7 5700G, libvirt VM (8 vCPU = 4 physical cores × SMT, 89.3 GiB RAM); GPU passthrough via xe driver
• Strata 5c8dd32 (v0.2.3 + max_tokens hotfix), engine 0.1.48, SYCL AOT build (bmg-g31)

Model provenance

• Repo: ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF, quant revision ed59f92082b1e93c0e96d60a8b11aab089b52f09, source/tokenizer revision de4b8e4d43b917e7706784d8bb445c9af86a3540 (HF LFS OIDs match the SHA-256s below)
• Files: shard 1 4c1eb2ceb4915e11…, shard 2 316b46f3a2dbd68c…, expert pack expert-profile-sycl.bin ddddf9b361dd1cb1… (identical to the run in PR [Benchmark] complete standard-draft.1 run: Intel Arc Pro B70 / Qwen3.8-Flash-Next GSQ-RCO IQ3_S #3799 — same model basis, clean version comparison)

Headline numbers (median, min–max in README)

Case	Prompt tok/s	Decode tok/s	TTFT s
short	n/a	50.4 (39.1–53.3)	0.90
prompt-4k	981.7	47.8 (40.0–47.9)	4.20
prompt-16k	1057.8	50.9 (50.4–51.5)	15.53

vs. the GPU-only-experts run of PR #3799 (engine 0.1.47): decode +26 % at 4k, +46 % at 16k; prompt rate −5…9 %. The host-arena path passed our 16k correctness gate twice before this run; on engine 0.1.44/0.1.45 the same switches produced token-0 output at ≥16k, so we re-gate after every engine update (details in the submitter addendum of the README).

Artifacts

• README.md sha256 9000bc2b0ddcf486…, results.json 178cdca55080c766…, outputs.jsonl bfc7abf4c29f2188…, engine-log-excerpt.txt 8a45489f11ce4e60… (full engine log for the run window, paths sanitized)
• Probe blind spots in the VM (driver version, GPU memory/power, CPU label, storage SMART) are listed in missing and covered in the submitter addendum.

Context

Follow-up to #3799 on the same machine/model; submitted per docs/COMMUNITY_BENCHMARKS.md submit-and-review flow. No community-table edit — reviewers decide.

[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.