ASUS has announced the UGen Intel Arc Pro B70 32GB, a graphics card it aims at local agentic AI. The card carries 32GB of GDDR6 with ECC and draws 310W, and four of them scale to 128GB of VRAM (press release).
One card is sized for most models under 27 billion parameters, and two cards ran 80B-class models
ASUS says a single card supports most models under 27 billion parameters and a context window of up to 128K tokens with Llama 3.1 8B, and that once VRAM runs out and data spills to host memory, inference performance can fall by up to 90%. Two cards give 64GB, and dual-card configurations ran 80B-class models at 43.4 tokens per second, though the model, the precision and the context length behind that figure are not given.
Support spans Intel’s own and open-source runtimes, and quantization needs no CUDA-based tool
ASUS says support covers Intel LLM Scaler, vLLM XPU, llama.cpp with SYCL and Vulkan, Ollama, OpenVINO, oneAPI, PyTorch XPU and ComfyUI. The verified models include Llama 3.1 and 3.2 from 8B to 70B, Qwen 3.x, DeepSeek-R1 Distill, Mistral and Mixtral, gpt-oss in 20B and 120B sizes, Whisper-large-v3 and InternVL3.5. LLM Scaler quantizes to INT4 or FP8 on the fly, which removes the need for an external CUDA-based quantization tool.
The price is called accessible but never given, and there is no ship date. Memory bandwidth and any single-card speed figure are also missing, so the one throughput number covers only the two-card case. Currently known ecosystem limitations do not affect near-term commercial deployment, in the release’s wording, without any examples.