Intel B70 -- Arc Pro Battlemage Lands with 32gb VRAM, 600+ GB/sec bandwidth, and more

TL;DR

Intel's Arc Pro B70 delivers 32GB VRAM and 620GB/s bandwidth as a serious local AI inference card, though software ecosystem gaps remain its biggest obstacle.

Key Points

• 1.The B70 is a workstation-class local AI card, not a gaming GPU. Built on XE2 Battlemage architecture with 32GB VRAM, ~250W TDP, and XMX tensor acceleration, it targets local LLM inference rather than competing with gaming cards like RTX or A770 successors.
• 2.620GB/s memory bandwidth and 32GB VRAM are the headline specs that matter. The reviewer emphasizes memory capacity over TOPS, ray tracing, or frame generation — at this price point and power envelope, it's uniquely positioned for running large local models.
• 3.Four-GPU inference benchmarks with Qwen 3.5 27B dense model show 369 average and 550 peak tokens/second. Running four B70s gives 128GB total VRAM, outperforming Strix Halo for inference at lower cost, with 200,000-token context windows and 50 concurrent requests.
• 4.Intel's LLM Scaler is a fork of vLLM, creating lag when new models like Qwen 3.5 drop. Support for Qwen 3.5 was added just before this video published — Intel also announced a vLLM partnership the same morning, which the reviewer called a positive development.
• 5.SR-IOV support is present on B70 with more virtual functions than B50, which was limited to two virtual functions despite earlier promises of four; Windows gaming drivers were not yet functional at launch time, with retail availability at B&H beginning the day after filming.