Reading the public Arc Pro B70 benchmark: XeStrata and SYCL

The XeStrata xe0.1.43.1 release compares Qwen3.8-Flash-Next IQ3_XXS on an Arc Pro B70 32 GB with the upstream v0.1.43 Intel SYCL build. The public generate results below use an approximately 4K-token input and 256-token output; prefill is input processing and decode is subsequent token generation.

This table reports the median for temperature-0 generate using the same IQ3_XXS model, an approximately 4K prompt, and 256 output tokens. The release alternated between the two runtimes for two runs; it does not include different context lengths or the ordinary chat path.

With a short prompt, a prefill difference may account for only a small part of the total wait; with a long answer, decode speed matters for longer. For a short question, measure time to first token⁵. For a longer coding request, also record generation speed and total completion time to see how these figures affect your work.

In xe0.1.43, 32K prefill for Qwen3.8-Flash-Next on an Arc Pro B70 fell back from the XMX matrix engine to the slower DP4a path; the release notes say it was about one-third slower. XeStrata xe0.1.43.1 fixes this and raises 32K IQ3_XXS prefill from 1,590 to 2,415 tok/s. This is a fix compared with an earlier XeStrata version, under conditions different from the 4K upstream comparison above.

The release also reports serving and everyday-use measurements separately from generate. For a 32K conversation, serving records prefill, the wait before a follow-up answer, and decode. Japanese chat and coding-agent tests go through serve/server.py and report first-token time and decode. For example, Japanese chat recorded 0.58 seconds to the first token and 62.7 tok/s decode. Each runtime⁶ generated its own response, with two alternating runs at temperature 0.7, so these figures should remain separate from the single generate test.

Read the public SYCL comparison in the release table; use xestrata --bench to record your own installation separately. It tests the installed IQ3_XXS for roughly 40 minutes. Stop the running XeStrata server first; add --bench-verbose if you need stage timings, GPU⁷ load, and temperatures.

Records are saved under records/bench-<time>/ in the program's data directory. Read summary.md and tables.md, then check record.json to confirm the hardware and engine paths. The default is two runs; increase it with an option such as --bench-rounds 3 for more repetitions. A mixed-GPU layer-split configuration measures everyday use only, not generate and serve, so check the scope before comparing tables.