User tests Llama.cpp on Intel iGPU: SYCL outperforms Vulkan

A user reports benchmarking Llama.cpp with an Intel Arrow Lake iGPU, finding that SYCL provides the best performance among GPU backends while Vulkan is largely ineffective.

• Vulkan support is described as broken or extremely slow, achieving only 0.5 tok/s on Qwen3.6 models and 4 tok/s on Gemma-4.
• SYCL runs Qwen3.6 35B models at 8-12 tok/s, with prefill speeds around 20 tok/s, though stability issues like freezing occur.
• CPU-only execution remains the most consistent option, delivering 14 tok/s for generation and 30-40 tok/s for prefill on the same hardware.

The results suggest that while SYCL can utilize the iGPU, it currently offers only marginal gains over CPU inference and suffers from reliability issues.