ASUS returns to Intel with a 32 GB Arc Pro B70 for local AI

ASUS previously sold an Intel DG1 card in 2021. As VideoCardz recalls, that passively cooled 4 GB model was mainly intended for prebuilt systems and required specific motherboard compatibility. The UGen Arc Pro B70 moves into a very different category.

This time, ASUS is adopting a professional card based on the Xe2 architecture. This return does not announce a gaming-focused Arc range: the manufacturer primarily presents an accelerator for AI applications and workstations.

The official specifications list 32 Xe cores and 256 XMX engines, alongside 32 GB of GDDR6 with ECC error correction. The 256-bit memory bus provides a stated bandwidth of 608 GB/s.

ASUS specifies a maximum clock speed of 2,800 MHz, up to 367 INT8 TOPS and board power of 310 W. The interface is PCIe 5.0 x16. The blower cooler and dual-slot design make it easier to install several cards in a suitable workstation.

VideoCardz also mentions a 12V-2×6 power connector, three DisplayPort 2.1 outputs and one HDMI 2.1b output.

ASUS highlights configurations with up to four cards, providing a combined 128 GB of VRAM. Its announced ecosystem includes Ollama, llama.cpp, vLLM, PyTorch and OpenVINO.

Understanding this proposition requires separating capacity from speed. More memory can help load larger models or retain more context. That alone does not predict generation speed: the model, its precision and the inference engine remain decisive.

Likewise, four 32 GB cards do not automatically behave like a single 128 GB card. The software must be able to distribute work and data across the GPUs. That needs checking before building a machine around this promise.

At the time of the announcement reported on 17 October, neither pricing nor a commercial availability schedule had been confirmed. A central piece of information is therefore still missing when assessing this offering against other local AI solutions.

Enlarge