Performance on Intel GPUs

[danielmayost]

I have an Intel ARC 140T iGPU (with XMX cores), llama.cpp has a Vulkan and sycl backend but for some reason it doesn't reach the performance that Intel can achieve like in ipex or openvino, why?

[NeoZhangJianyu]

IPEX and OpenVINO are commercial software.
There are more engineers to optimize the performance on Intel GPU.
New technology is applied by them.
The important factor is their optimization to be based on different hardware feature/platform.

If you want to get the ultimate performance on Intel GPU, please use them. (IPEX is replaced by Pytorch XPU; for service, recommend vLLM/sglang for Intel).
But they are heaver than llama.cpp.