Intel Arc Pro B70 hits 2.2k tps

Intel's Xe2 architecture is finally showing its teeth in AI, but the "Intel Tax" of messy software setup is still very real for early adopters.

• –Ingestion performance (2,200 tps) is highly competitive, but generation speeds (~10 tps) suggest a massive bottleneck in the current vLLM/SYCL implementations for Battlemage.
• –vLLM with the `intel/llm-scaler-vllm` docker image is the only reliable path for high performance, as native `llama.cpp` backends (OpenVINO, Vulkan) are still throwing critical tensor errors.
• –The 32GB VRAM on the Pro variants makes 27B+ models viable locally, but the requirement for Ubuntu 26.04 or Fedora 43 for proper kernel support is a steep barrier for casual users.
• –Sub-par generation speeds relative to ingestion highlight that Intel's XMX engines are underutilized in current autoregressive sampling loops compared to NVIDIA's Tensor Cores.