Intel Arc Pro B70 - Benchmarks & Performance Data
A collection of measured benchmark results for the Intel Arc Pro B70 GPU covering LLM inference speed and power draw, video generation, and dual-card scaling, from real single- and dual-card runs.
These benchmarks test llama.cpp with its SYCL and Vulkan backends and vLLM's XPU support, plus several video-generation pipelines, on a single Intel Arc Pro B70 card with 32GB of memory and on a dual-card configuration with 64GB total. The project describes public data on this card as scarce, so it collects the numbers, the rough edges hit along the way, and the fixes found for them.
A results table reports tokens per second for prompt processing and generation, average power draw and tokens per joule across dense and mixture-of-experts models ranging from about 1.5 billion to 80 billion parameters, all measured on one specific, commit-pinned llama.cpp build with power telemetry captured during every run. It is licensed MIT.