ASUS has announced the UGen Arc Pro B70, its first Intel Arc card aimed at local artificial intelligence inference. The UGen Arc Pro B70 features 32GB of ECC GDDR6 memory and positions itself as an economical option to run AI models without relying on the cloud, aligned with the company's Agent Computers vision.
Main specifications of the UGen Arc Pro B70
The card uses Intel's Xe2 architecture on the BMG-G31 chip, with 32 Xe-cores and 256 XMX engines. According to ASUS, a single UGen Arc Pro B70 supports most models under 27 billion parameters and enables context windows of up to 128K tokens on models such as Llama 3.1 8B. Dual configurations reach 64GB and four cards reach 128GB of combined VRAM.
Power consumption is 310W (above Intel's 230W reference), with a dual-slot blower cooler, 12V-2x6 connector, PCIe 5.0 x16 interface and outputs of three DisplayPort 2.1 plus HDMI 2.1b. Software support includes llama.cpp (SYCL and Vulkan), Ollama, vLLM XPU, OpenVINO, PyTorch XPU and ComfyUI.
Why the UGen Arc Pro B70 matters for local AI
Demand for local and multi-agent AI is growing rapidly. High-VRAM cards at a more accessible cost help prosumers, studios and enterprise teams keep sensitive data on-premises, avoiding performance drops of up to 90% when VRAM runs out and the system offloads to main memory.
There is still no official global price in the announcement, but the focus is memory capacity per cost for inference workloads. More details in the official ASUS press release.
Image credit: ASUS — Source: ASUS Pressroom
By GeekikiBot
