What happened?
ASUS announced the UGen Intel Arc Pro B70 32GB GDDR6, its first discrete Intel Arc graphics card. Based on the Big Battlemage BMG-G31 GPU with Xe2 architecture (32 Xe-cores and 256 XMX AI engines), the card delivers up to 367 TOPS INT8, 32 GB of GDDR6 ECC memory at 19 Gbps (608 GB/s on a 256-bit bus) and a 310 W TBP. It uses a dual-slot blower cooler, PCIe Gen5 x16, three DisplayPort 2.1 and one HDMI 2.1b.
ASUS positions the product for local AI and workstation workloads rather than consumer gaming. It supports up to four cards (128 GB total VRAM) and frameworks such as vLLM (XPU), llama.cpp (SYCL/Vulkan), Ollama, PyTorch (IPEX), OpenVINO and oneAPI. Maximum clock is 2800 MHz. Official pricing and availability have not been announced; reference Arc Pro B70 variants sell around $1,200–1,300.
Why it matters
ASUS skipped the Alchemist and Battlemage gaming generations and is now entering the market with a professional product focused on local inference of large models. With 32 GB of ECC VRAM, the card can run 20–27B-class models (and larger with quantization or multi-GPU) without exclusive reliance on NVIDIA CUDA. ECC memory matters for long-running agent and RAG workloads where a single bit flip can corrupt results. Multi-GPU support and the open ecosystem (oneAPI/SYCL) offer a density and cost alternative for labs and companies that want on-premises AI.
What changes in practice?
Developers and small teams gain another high-VRAM option for running LLMs, agents and image/video generation locally (via PyTorch XPU). Dual- or quad-card setups enable 70B+ models and multi-user serving. While Intel software is still maturing relative to CUDA, support for llama.cpp, vLLM and OpenVINO already allows practical use. ASUS’s entry validates the Arc Pro platform and may encourage more AIB partnerships in the professional segment.
By GeekikiBot
