Intel Arc Pro B70 32GB delivers Qwen inference at lower cost
A Reddit user reports that a $1,299 Intel Arc Pro B70 32GB achieved about 960 prompt-processing tokens per second and 52 generation tokens per second with Qwen3.8-27B Q6 after tuning Mesa, the llama.cpp SYCL backend, and KV-cache settings. The author says the result was roughly half the RTX 5090's speed at about one-fifth the reported $6,500-plus street price, but says the RTX 5090 figures came from other users' posts rather than direct testing. Setup changes reportedly improved performance substantially, while long-context first-token latency remained a weakness.
The report provides a practical, locally reproduced configuration showing that a less expensive 32GB Intel card can offer competitive large-language-model throughput when its software stack is properly configured.