ASUS Puts a 32GB AI Inference Card on the Desktop and Calls It the Building Block of an Agent Computer
The UGen Intel Arc Pro B70 pairs 32GB of ECC GDDR6 with an inference-tuned architecture, scales to 128GB across four cards, and ASUS says a single card covers most models under 27 billion parameters. The memory argument is the center of ASUS's case.
ASUS says it has announced the UGen Intel Arc Pro B70 32GB, a graphics card it describes as an economical option for the fast-growing market in local agentic AI. In a press release on its own pressroom, the company frames the card as an answer to what it calls the market's most pressing demand: ample VRAM at an accessible price. ASUS says the card pairs 32GB of GDDR6 ECC memory with an inference-optimized architecture, and that the combination is intended for running AI locally, privately and economically as large language models and multi-agent workflows move off the cloud and onto desktops and workstations.
The memory argument is the center of ASUS's case. The company says modern AI workloads are memory-hungry, that long-context reasoning dramatically inflates KV-cache usage, and that multi-agent deployments run several models at once and multiply VRAM requirements. According to ASUS, once VRAM runs short the data offloads to host memory and inference performance can drop by up to 90 percent. The company says the card's 32GB of ECC-protected GDDR6 can be enough to keep demanding models entirely local, and that a single UGen B70 supports most models under 27 billion parameters with a context window of up to 128K tokens on Llama 3.1 8B. ASUS says the configuration scales to 64GB with two cards and 128GB with four.
ASUS also lists the software stack it says supports the card, naming Intel LLM Scaler and vLLM XPU, llama.cpp through SYCL and Vulkan, Ollama, Intel OpenVINO and oneAPI, PyTorch XPU, and ComfyUI for image generation. The company says verified models span Llama 3.1 and 3.2 from 8B to 70B, Qwen 3.x, DeepSeek-R1 Distill from 7B to 70B, Mistral and Mixtral, OpenAI gpt-oss in 20B and 120B versions, Whisper-large-v3, and InternVL3.5. ASUS says dual-card configurations have been proven to run 80B-class models at 43.4 tokens per second, and that LLM Scaler's built-in INT4 and FP8 online quantization removes the dependency on external CUDA-based quantization tools.
The release splits the audience into three groups. ASUS says prosumers and independent AI developers get past the resource ceiling that mainstream 16GB cards impose on larger models, that small businesses and creative studios can use dual-card 64GB setups to support larger deployments while keeping sensitive data in-house, and that enterprise departmental AI teams can use multi-card 128GB systems to cut long-term costs against cloud APIs while meeting data-compliance requirements in fields such as law, medicine, government and finance. ASUS also states the card draws 310W, which it presents as part of the cost equation for local deployment.
ASUS says this scalability is what makes the UGen B70 the building block of what it calls an agent computer, a system designed from the ground up to run AI agents locally, and it points readers to its Agent Computer page for the wider vision covering cloud, local and hybrid AI deployments. The company describes itself in the release as a technology leader delivering AI solutions across infrastructure, physical AI, AI PC and devices, and notes the UGen B70 announcement carries a Taipei dateline of October 23, 2026.
Source: ASUS (https://press.asus.com/news/press-releases/asus-ugen-intel-arc-pro-b70-32gb-ai-accelerator/).
Source: ASUS (https://press.asus.com/news/press-releases/asus-ugen-intel-arc-pro-b70-32gb-ai-accelerator/).
InteliToys is editorial. We do not sell ASUS ASUS UGen Intel Arc Pro B70 32GB. Watch the source tape in the branded lightbox below, then argue about it in Gear Junkies.