Add Arc Pro B70 results: Qwen3.6-27B and Qwen3.8-27B

[shacortes]

Six single-GPU results on one Intel Arc Pro B70, single request at a time.

model	runtime	quant	backend	decode tok/s	prefill tok/s
qwen3-6-27b	llama.cpp	q4_k_m	SYCL	26.64	312.6
qwen3-6-27b	llama.cpp	q4_k_m	Vulkan	11.49	679.7
qwen3-6-27b	vLLM	int4	XPU	27.96	—
qwen3-8-27b	llama.cpp	q4_k_m	SYCL	20.35	288.6
qwen3-8-27b	llama.cpp	q4_k_m	Vulkan	14.81	679.6
qwen3-8-27b	vLLM	int4	XPU	27.01	—

Provenance

Both runtimes are stock upstream checkouts with clean trees — no fork, no local patches, so customRuntime is unset on all six.

• llama.cpp b11026 (b49650adb), remote ggml-org/llama.cpp, built for SYCL and Vulkan
• vLLM v0.29.0 (98dff2a81), remote vllm-project/vllm, XPU device, third-party plugins disabled

Method

This was run headless, to reduce GPU usage due to graphics and only keep the processes needed to run inference.

llama.cpp numbers come from llama-bench at -p 512 -n 256 -r 5 -ngl 99 -fa 1; decodeTps is tg256 and promptTps is pp512, each averaged over 5 reps.

vLLM numbers come from vllm bench serve (random dataset, 8 prompts × 64 in / 256 out, --max-concurrency 1). decodeTps is 1000/mean_tpot_ms so it measures the same thing as llama.cpp tg256 rather than throughput over total wall time, which includes load and queueing — both figures are in the per-result notes.

promptTps is omitted for the two vLLM rows rather than estimated. vllm bench serve reports no prefill throughput, and the engine's Avg prompt throughput log line averages over a 10s window, which at 512 total prompt tokens reads 6.4 tok/s — diluted by decode and idle time, so not a prefill figure. Would need a re-run against vllm:request_prefill_time_seconds at a comparable prompt length.

[vercel]

@shacortes is attempting to deploy a commit to the Community Labs Team on Vercel.

A member of the Team first needs to authorize it.

[github-actions]

- results/shacortes/qwen3-6-27b-llamacpp-q4_k_m-sycl.json: validated
- results/shacortes/qwen3-6-27b-llamacpp-q4_k_m-vulkan.json: validated
- results/shacortes/qwen3-6-27b-vllm-int4.json: validated
- results/shacortes/qwen3-8-27b-llamacpp-q4_k_m-sycl.json: validated
- results/shacortes/qwen3-8-27b-llamacpp-q4_k_m-vulkan.json: validated
- results/shacortes/qwen3-8-27b-vllm-int4.json: validated
Ready for merge: the author has a linked GitHub account and the database accepted the results in a rolled-back validation transaction. Merging submits them automatically; no web form is needed.

Site / sign up · Workflow details and retry

[shacortes]

A bit doubtful on the Qwen3.6 27B llama.cpp Vulkan build's results.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
intelinside	Ready	Preview	Sep 17, 2026 7:41pm UTC