XeCores, an Intel Arc field guide for local AI

Field notes

XeCores structures its material around repeatable inference conditions rather than undifferentiated GPU performance claims. Its benchmark rules distinguish prompt processing from token generation and require power caps and configured draw to accompany measurements. The site describes six validated model families across llama.cpp SYCL and vLLM XPU, plus a ComfyUI VideoGen setup using dual B70 cards, with a single-card CPU fallback.