ASUS has announced the UGen Intel Arc Pro B70 32GB, an AI accelerator card built for local agentic AI workloads on desktops and workstations – it features 32GB of ECC GDDR6 memory with an inference-optimized architecture, aiming to bring cloud-scale AI capabilities in-house without the recurring costs of cloud APIs.
ASUS UGen Intel Arc Pro B70 32GB
The card is built on Intel’s Xe2 architecture, featuring 32 Xe-Cores and 256 XMX AI engines. It delivers up to 367 TOPS of INT8 AI performance, with a maximum GPU clock of up to 2800MHz. Memory bandwidth reaches up to 608GB/s over a 256-bit interface, and the card connects via PCIe 5.0 x16.
A single card supports most large language models under 27 billion parameters and enables a context window of up to 128K tokens when running Llama 3.1 8B. Multi-card scaling allows the system to fit larger models: dual-card configurations provide 64GB of combined VRAM, and four-card setups reach 128GB, with dual-card systems already proven to run 80B-class models at 43.4 tokens per second.
The ASUS UGen Arc Pro B70 is supported by a broad software stack that includes Intel LLM Scaler, vLLM XPU, llama.cpp (SYCL and Vulkan), Ollama, Intel OpenVINO and oneAPI, PyTorch XPU, and ComfyUI for image generation. LLM Scaler includes built-in INT4/FP8 online quantization, removing the need for external CUDA-based quantization tools. Verified compatible models include Llama 3.1/3.2 (8B to 70B), Qwen 3.x, DeepSeek-R1 Distill (7B to 70B), Mistral and Mixtral, OpenAI gpt-oss (20B and 120B), Whisper-large-v3, and InternVL3.5.
The ASUS UGen Intel Arc Pro B70 32GB will be available from early December 2026 at a recommended retail price of RM8,459.
Pokdepinion: Most businesses is probably going to happily pay for cloud subscriptions, but in cases where data is sensitive, this might be an option to look for.
