mesa PR with 37-130% llama.cpp pp perf gain for vulkan on Linux on Intel Xe2

Posted by TheBlueMatt@reddit | LocalLLaMA | View on Reddit | 4 comments

[-]

hainesk@reddit

I've been using oneapi with llamacpp. How would this compare? Because llamacpp with vulkan was pretty slow for my 258v and using oneapi with syscl has been much better

[-]

seasea128@reddit

I've been using llamacpp with vulkan on my lunar lake laptop and it was faster than syscl for me. this is on the docker image provided by llamacpp, so could be because of different version of oneapi?

[-]

This_Maintenance_834@reddit

looks like I returned my intel B70 too early.

[-]

There's likely more to come, too. The patch at https://gitlab.freedesktop.org/mesa/mesa/-/work_items/15312#note_3443232 is a bit harder to upstream but it shows another 8% perf gain on BF16 models (on top of the 2.3x in the upstream PR).