Hi, I recently tried to get llama.cpp with SYCL running on an Arrow Lake system but gave up halfway through since Vulkan is just way easier to set up. But, the pp/tg I'm getting on Vulkan w/ Arc 130T is disgustingly bad - 100 tokens/s for pp256 and less than 4 for tg64 with Gemma 4 E4B, worse than any newish CPU I've tried previously. Do these get any better with SYCL, or what else am I supposed to use with Intel iGPUs? I'm unironically getting better tg speed on Zen 4 iGPUs with vulkan lmao
[link] [comments]