whisper-bench SYCL faster than BLAS faster than VULKAN on Intel CPU+iGPU: i5-11400 for large-v3. Why?

[arkanoid87]

I'm new to whisper.cpp and trying to find the fastest setup to run it on Intel CPU + iGPU

By following the project readme I managed to compile (git master branch) and run whisper-bench on 3 backends (+1 failure)

• SYCL: total time = 72231.18 ms
• BLAS: total time = 142385.89 ms
• VULKAN: total time = 222731.31 ms
• OPENVINO: no iGPU detected (from openvino import Core; Core().available_devices)

bench_vulkan.txt
bench_sycl.txt
bench_blas.txt

I was expecting OPENVINO to be the fastest, VULKAN the second place, but actually I got very different results: SYCL seems 3x faster than VULKAN. I'd like to ask you if you see any error in my benchmark, and if you consider my results reliable. Should I stick with SYCL? Should I fix the setup for a different -more performant- backend?

Thanks