llama.cpp Vulkan stumbles on Arrow Lake

This looks more like backend maturity and memory-bandwidth limits than a hardware surprise. Intel iGPUs are supported, but the post shows why Vulkan still feels like the fallback path rather than the preferred Intel stack.

• –Intel’s llama.cpp docs position SYCL as the primary backend for Intel GPUs, and explicitly list Arrow Lake’s built-in Arc graphics as supported.
• –The numbers fit a familiar pattern: prompt processing can look acceptable while token generation falls apart on integrated graphics.
• –OpenVINO is the other Intel-specific lane worth watching; Vulkan is easier to set up, but not the obvious choice for throughput.
• –For users who want predictable local LLM performance today, a tuned CPU build or a discrete GPU still looks safer than betting on an Intel iGPU backend.