Intel Arc Pro B70 hits 282 t/s prompt eval

The report confirms that llama.cpp’s SYCL backend is now mature enough for production-grade speeds, significantly outperforming Vulkan on Battlemage hardware. Successful deployment on a PCIe 3.0 system proves the architecture's resilience to older bandwidth standards, extending the life of legacy workstations. Furthermore, performance spikes in prompt evaluation suggest that Intel's driver-level optimizations for Flash Attention are delivering competitive throughput. At $949, the card enables running large models like Qwen 3.6 35B with massive 130k context windows entirely in VRAM, effectively undercutting the "Nvidia tax" for local inference.