Builds / Intel Arc B390 iGPU (Core Ultra X7 358H, Panther Lake, share
Intel Arc B390 iGPU (Core Ultra X7 358H, Panther Lake): Qwen3.6 35B-A3B Q4_K_M at 30.9 tok/s on Vulkan vs 28.3 on SYCL (llama.cpp)
Last updated:
- GPU(s)
- Intel Arc B390 iGPU (Core Ultra X7 358H, Panther Lake, shared memory)
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 23.7
- CPU
- Intel Core Ultra X7 358H (16 cores)
Amazon · eBay (used) · Newegg - Motherboard
- laptop (not stated)
Amazon · eBay (used) · Newegg - RAM
- shared system memory (23.7 GiB visible to Vulkan, 28.8 GiB to SYCL)
Amazon · eBay (used) · Newegg - PSU
- not stated
- Other
- CachyOS, Linux 7.3.1, Mesa 26.2.5 (ANV), Level Zero 1.32.0, oneAPI 2026.1; llama.cpp 0.9.0-dev commit 9c9074e14, Vulkan build with GCC 16.2.1, SYCL build with icx (F16 off, SYCL graph off); llama-bench pp512/tg128, 16 threads, 3 reps
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.6 35B-A3B | Unsloth UD-Q4_K_M with MTP head (21.1 GiB) | llama.cpp 0.9.0-dev 9c9074e14 (Vulkan) | 30.9 | 758 @ 512-token prompt est. TTFT @8k ≈ 10.8 s (calculated from measured prefill) | context: llama-bench tg128📏 llama-bench -ngl 99 -dev Vulkan0 -t 16 -r 3: tg128 30.9 ± 0.1; 32 layers on GPU 18.8, CPU only (ngl 0) 13.1 |
| Qwen3.6 35B-A3B | Unsloth UD-Q4_K_M with MTP head (21.1 GiB) | llama.cpp 0.9.0-dev 9c9074e14 (SYCL, oneAPI 2026.1) | 28.3 | 439 @ 512-token prompt est. TTFT @8k ≈ 18.7 s (calculated from measured prefill) | context: llama-bench tg128📏 llama-bench -ngl 99 -dev SYCL0 -t 16 -r 3: tg128 28.3 ± 0.3; 32 layers on GPU 24.1; SYCL build on CPU only 20.7 (vs 13.1 for the standard CPU backend) |
Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 ÷ prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.
Owner's own llama-bench runs, published 5 Nov 2026 on grigio.org. The article title says Qwen3-35B-A3B but the commands load Qwen3.6-35B-A3B-MTP-UD-Q4_K_M.gguf, so it is listed as Qwen3.6 35B-A3B; MTP drafting was not used. The Arc B390 is an integrated GPU sharing system RAM; Vulkan won clearly on prompt processing (758 vs 439) and slightly on generation. Synthetic single-user numbers.
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.