Intel Arc Pro B65 benchmarks

Measured local-LLM performance on an Intel Arc Pro B65 32 GB: llama.cpp Vulkan and SYCL, three 4-bit GGUF models, throughput, server latency, board power, cache behavior, and small quality checks. The main experiments ran on October 12–13, 2026, with registry checks on October 14.

• Article: the settings I'd use — backend comparisons, per-model results with standard deviations, practical commands, and an experiment log.
• Recommended settings — the final recipe and its supporting measurements.
• Findings and raw-run index — each experiment, its findings, and run identifiers; dry runs are identified separately and excluded from conclusions.
• Frozen article evidence — the evidence snapshot cited by the article. Article citations use the full commit ID so they continue to identify the same files after updates.

Tested stack: llama.cpp build 11574 (8bb3cf0d2), oneAPI 2026.0, compute-runtime 26.31.39395.13, Mesa 26.2.2, kernel 7.2.5-3-omarchy. B65 only; an RTX 3080 drove the display and was excluded from inference. Power readings measure the GPU board, not the whole PC.

The repository includes scripts, raw JSONL/CSV/logs, model-file SHA-256 manifests, findings, and share-card images. Model weights and vendor/build directories are excluded. Recorded local paths and device identifiers in the raw files describe the test machine; adjust paths and device selection when reproducing a run.

Layout

Path	What
findings/	Per-phase findings (verified against the raw data) + an index
run.sh, report.py, powerwin.py, models.conf	Phase 0 kit (prefill + decode sweep); raw/ and charts/ are its runs; results.csv is the merged table
b65-phase05/	Phase 0.5 kit (flash attention off vs on) and its raw/ runs
b65-phase1/	Phase 1 kit (context depth, llama-server, loaded idle), the Gemma checkpoint rerun script, Forge's HANDOFF.md and its raw/ runs
b65-phase2/	Phase 2 kit (Vulkan vs SYCL), local fixes listed in its README, and its raw/ runs
dashboard/	Live tailnet dashboard + share cards (server.py, one page per phase, share.html)
holds.conf	Results withheld from share cards (currently none)
dist/	Release tarballs (not committed; regenerated from the raw runs)

Raw run folders are never edited after a run; new runs always go to new folders.

Phase 0 Vulkan setup (no sudo needed)

The commands below reproduce the initial Vulkan baseline. For the final SYCL setup, see the article's build instructions and the Phase 2 kit.

Everything lives in the project dir; vendor/, .venv/ are not part of the results.

uv venv .venv && uv pip install --python .venv/bin/python cmake ninja matplotlib
# headers matching the installed vulkan-icd-loader (1.4.357), installed to vendor/prefix
git clone --depth 1 -b v1.4.357 https://github.com/KhronosGroup/Vulkan-Headers vendor/Vulkan-Headers
git clone --depth 1 -b vulkan-sdk-1.4.357 https://github.com/KhronosGroup/SPIRV-Headers vendor/SPIRV-Headers
# (cmake -B build + cmake --install build -DCMAKE_INSTALL_PREFIX=vendor/prefix for each)
git clone https://github.com/ggml-org/llama.cpp vendor/llama.cpp && git -C vendor/llama.cpp checkout b11574
cmake -S vendor/llama.cpp -B vendor/llama.cpp/build -G Ninja -DCMAKE_BUILD_TYPE=Release \
  -DGGML_VULKAN=ON -DGGML_CUDA=OFF -DGGML_NATIVE=ON -DLLAMA_CURL=OFF \
  -DCMAKE_PREFIX_PATH=$PWD/vendor/prefix -DVulkan_INCLUDE_DIR=$PWD/vendor/prefix/include
cmake --build vendor/llama.cpp/build -j --target llama-bench llama-completion

Find the B65's Vulkan index with vendor/llama.cpp/build/bin/llama-bench --list-devices (on this machine: 1; 0 is the 3080). Unload LM Studio models first (lms unload --all); run.sh refuses to start with < 28 GB free on the B65.

Run

B65_VK_INDEX=1 DRY=1 ./run.sh   # one model, pp512 + spot-check
B65_VK_INDEX=1 ./run.sh         # full run, ~1 h

Each test: llama-bench's built-in warmup, then 5 repeats (mean +/- sd), flash attention on, all layers on GPU. Tests: prefill 512 / 4096 / 16384 tokens; decode 128 tokens at 0 / 4K / 16K context depth. Each model also gets a greedy 96-token spot-check generation, to catch fast-but-garbage output.

Power

B65 board power from the xe hwmon card energy counter (not wall power), sampled at 10 Hz during each test. board_w_avg averages over the timed reps only (the trailing sum(samples_ns) window before llama-bench exits), which excludes model load and warmup. It is left blank for -d depth tests, because llama-bench runs an untimed depth prefill inside every rep and the window would mix prefill and decode power. invocation_w_avg (whole process including load) is kept for transparency but is not a headline number.

Output

• raw/<run_id>/: env.txt, llama-bench JSONL + logs, power traces, energy.csv, spot-checks, model sha256s (never edit)
• results.csv: one row per model x test (failures recorded with status), appended every run
• charts/<run_id>-pp.png, -tg.png: 16:9 charts with versions in the footer

Machine notes (logged in env.txt)

• The B65 links at PCIe 4.0 x4 via the X570 chipset (card supports x16). Arc's on-card switch makes the endpoint itself report 2.5 GT/s x1; the real link is read from the upstream bridge.
• ReBAR is on (BAR2 = 32768 MiB).
• With nothing loaded, Vulkan reports ~29.3 GB of 32 GB free on the B65 (driver-reserved; no process holds it).

Live dashboard (tailnet only)

The dashboard is optional and runs on your own machine or tailnet. Public benchmark evidence is available directly in this repository; no dashboard login or tailnet access is needed to read it.

Example: https://benchmark-host.example.ts.net:8445/ serves live board power/temps/fan, current test and rep, results table as tests finish, spot-checks, env. Read-only; it never touches the GPU.

systemd-run --user --unit=b65-dashboard --working-directory=$PWD .venv/bin/python dashboard/server.py --port 8765
tailscale serve --bg --https=8445 http://127.0.0.1:8765      # off: tailscale serve --https=8445 off

Start long runs as a user service so they survive the terminal/agent session that launched them:

systemd-run --user --unit=b65-bench-run --working-directory=$PWD -E B65_VK_INDEX=1 -E PATH=$PATH -E HOME=$HOME \
  bash -c './run.sh > full-run.log 2>&1'
systemctl --user stop b65-bench-run    # abort

Share cards

/share has one tab per phase (/share#p0, #p05, #p1, #p2; the sidebar's Share cards link opens the current phase's tab). Each tab shows that phase's run status, then Post now (no measured numbers) and Results (unlock when that phase's run finishes). Only the open tab is drawn. A new card joins a tab through its group; a new phase needs an entry in SHARE_PHASES in share.html.

On your dashboard, /share serves social-size PNG cards (16:9 for X/LinkedIn, 4:5 for Instagram/Threads) with Download / Copy image / Share and a suggested caption for each.

• Teasers (live progress, proof: the model's own spot-check output, redacted results, lineup, methodology, test rig) never contain measured values. While a run is in progress /api/share sends only per-test status (done / running / pending; failures are also reported as done) and a unitless power shape, so the page has nothing to leak. Redaction blocks are all the same width.
• Results cards (findings, per-model build card / spec sheet, decode, prefill, tokens per joule, headline) unlock only when the latest run has a done marker and every planned test has finished.
• Phase 0.5 cards (own section): teasers (preview, live sweep grid, poll, exact model versions) and results (% change with flash attention off, settings from here on, Gemma before/after) that unlock when the sweep finishes. Holds apply to both flash-attention settings; the Gemma card stays locked until its prefill is released.
• Phase 1 cards (own section): teasers (preview, live grid, "time to first word" poll) and results (context-fit grid, time to first token, key findings) that unlock when the Phase 1 run completes. OOMs report as "done" until then.
• Holds: holds.conf lists results that stay unpublished even after a run completes (model|test|reason). The server strips them from /api/share, and cards show "held". Delete a line to release it.
• One card at its exact size: share?solo=1&card=<live|proof|p05|p05-live|p05-bets|p05-prov|p05-findings|p05-delta|p05-decode|p05-settings|p05-gemma|p1-preview|p1-live|p1-guess|p1-context|p1-ttft|p1-findings|redacted|lineup|method|rig|findings|build|decode|prefill|efficiency|headline>&fmt=<wide|tall>[&model=<name>][&handle=@you][&price=<what you paid>]
• Captions come in variants (post / short take / reply), written short, first person, ending in a question, with a 280-character counter. Card price is typed in by hand on the page and only appears once set.

Phase dashboards

Every page has a left sidebar (dashboard/nav.js, statuses from /api/nav) linking all phases, share cards, findings and files, with a live status dot per phase. To add a phase, add it to the PHASES list in nav.js and to nav_status() in server.py.

• /p05: Phase 0.5 flash-attention sweep (off vs on, with Phase 0 reference and provenance).
• /p1: Phase 1 (decode by context depth with OOMs, llama-server TTFT, loaded idle power + VRAM, drift check vs Phase 0.5).
• /p2: Phase 2, Vulkan (FA off) vs SYCL (FA off/on) per model and test, SYCL build state, decode verdict so far.
• /p1g: Phase 1 follow-up, Gemma llama-server rerun (b65-phase1/gemma_server_rerun.sh → raw/p1-swa-*): --ctx-checkpoints 0 / --swa-full / defaults vs the Phase 1 baseline.

Files

On your dashboard, /files/ is a read-only browser/download route for findings/, dist/, charts/, raw/, the phase kits, results.csv and this README (nothing else in the project or on the machine). From the Mac you can also use scp user@benchmark-host:/path/to/b65-benchmarks/dist/<file> . over your configured SSH connection. Findings moved from dist/*-findings.md to findings/phase{0,05,1}.md on 2026-10-13.