Awesome Arc Pro B70

Lab notes from our B70 test queue, published as written (2026-10-15). Raw logs are not included. Back to the performance guide.

002 result (TRIMMED per user: <=20 GPU min): Mesa 26.2.5 vs 25.3.6, llama.cpp Vulkan, B70, 2026-10-15 — PASS (stack), master tbo NOT RUN

Trim: Mesa 26 only (Mesa 25.3.6 numbers = RUN.md table, measured 2026-10-15 on the same builds), d=0 only, 35B-A3B pp512/tg128 and 27B tg128, r=3; test-backend-ops MUL_MAT + FLASH_ATTN_EXT. Raw: RESULT.txt, results/. Driver line confirms “Mesa 26.2.5”.

model / test	master Mesa 25	master Mesa 26	stack Mesa 25	stack Mesa 26	SYCL master
35B-A3B tg128 d=0	39.0	86.14 ±0.94 (+121%)	64.3	87.32 ±0.34 (+36%)	97.7
35B-A3B pp512 d=0	1854	1637 ±60 (−12%)	1931	1737 ±25 (−10%)	1301
27B tg128 d=0	13.2	22.85 (+73%)	22.7	25.26 ±0.14 (+11%)	21.9

test-backend-ops on Mesa 26: stack MUL_MAT 2279/2279, FLASH_ATTN_EXT 5321/5321 (PASS). master: not run — its build dir has no test-backend-ops binary (the “FAIL=0” lines for master are empty output, not a pass).

Reading:

• Mesa 26 fixes most of master’s Vulkan decode problem by itself (35B 39 → 86, 27B 13 → 23). The hub/doctor “use SYCL” advice needs a “Mesa >= 26” qualifier: on Mesa 26, Vulkan master 35B-A3B decode is 88% of SYCL (86.1 vs 97.7) and beats SYCL on prefill.
• On top of Mesa 26 the stack still adds +1.4% (35B tg) and +10.5% (27B tg), so the 32-bit-load change is not just a Mesa-25 workaround for the dense model; on the MoE it mostly washes out.
• Mesa 26 costs ~10–12% pp512 on 35B-A3B for both arms (worth a look before recommending it for prefill-heavy use).
• No xe faults during the run.

This site is open source. Improve this page.