GUNNIR Arc Pro B70 Dual-Card Liquid-Cooled Workstation and Mini Workstation Make Offline Debut

【Core Events and Technical Overview】

At the recent Intel Technology Innovation and Industry Ecosystem Conference, GUNNIR showcased its dual-card liquid-cooled workstation and mini workstation solutions based on the Arc Pro B70 graphics card offline for the first time. Both devices are powered by the Intel Core Ultra 200HX "Arrow Lake HX" processor platform. Through the ingenious design of liquid cooling and vapor chamber heat dissipation solutions, they achieve a dual breakthrough in memory capacity and AI computing power density within an extremely limited physical volume. The core goal of this exhibition is to provide hardware-level support for local AI application scenarios, marking the accelerated evolution of professional-grade AI computing power towards desktop and miniaturized forms.

In addition to the high-spec B70 series workstations, GUNNIR simultaneously exhibited MXM form factor graphics cards based on the Arc Pro B50 and B65. MXM modular design has historically been favored in compact industrial PCs and mobile workstations; introducing it into the AI computing field further enriches the forms of edge-side AI deployment. Overall, GUNNIR's exhibited product matrix covers everything from high-power dual-card liquid cooling to highly integrated mini workstations, as well as flexible pluggable MXM modules, building a local AI computing infrastructure ecosystem tailored for different computing power needs and space constraints, filling the computing power gap between consumer-grade graphics cards and cloud data centers.

【Technical Principles and Core Breakthroughs】

From the underlying technical architecture, the Arc Pro B series graphics cards are built on Intel's Xe2 graphics architecture, which has been deeply optimized for AI computing. Its core XMX (Xe Matrix Extensions) engine is specifically designed to accelerate matrix operations in deep learning, efficiently handling dense GEMM operations during large model inference. Compared to traditional SIMD computing units, the XMX engine achieves a significant increase in throughput under FP16 and INT8 precision. Furthermore, paired with Intel's oneAPI software stack and OpenVINO toolkit, developers can bypass the proprietary CUDA ecosystem to achieve seamless integration and underlying operator-level optimization for mainstream deep learning frameworks like PyTorch and TensorFlow.

At the engineering architecture level, the core breakthrough of the dual-card liquid-cooled workstation lies in thermal management and multi-card interconnect design. Placing two high-power professional graphics cards in a compact chassis makes heat dissipation the biggest engineering challenge. GUNNIR's solution combining a custom liquid cooling loop and vapor chamber not only covers the GPU core but also accommodates the memory chips and power delivery modules, ensuring temperature stability during prolonged, high-load AI inference. Meanwhile, leveraging the high-bandwidth PCIe lanes provided by the Core Ultra 200HX platform, the dual cards can achieve efficient data parallel communication. Although it does not currently support memory pooling similar to NVLink, the dual-card system can still significantly reduce latency in large model inference and improve overall throughput through software-level tensor parallelism and pipeline parallelism technologies.

【Industry Background and Competitive Landscape】

Currently, the global AI computing power market is in a critical transition period of expanding from the cloud to the edge and device sides. As the parameter scale and performance of open-source large models like Llama 3 and Qwen continue to rise, the demand from enterprises and developers for localized AI computing power with data privacy protection capabilities is experiencing explosive growth. Against this backdrop, while NVIDIA's RTX Ada architecture professional graphics cards dominate the market, their high prices limit adoption among small and medium-sized enterprises and research institutions. The launch of GUNNIR's Arc Pro B70 workstation targets this market gap, attempting to break NVIDIA's monopoly in the local AI workstation sector with higher memory capacity cost-effectiveness and an open software ecosystem.

Regarding the competitive landscape, AMD recently also introduced large-memory graphics cards for workstations, such as the Radeon PRO W7900, attempting to attract AI developers through the ROCm ecosystem. In contrast, the advantage of Intel's Xe architecture lies in its deep coupling with the x86 CPU ecosystem and the cross-platform abstraction capability of oneAPI. As a core partner of Intel, GUNNIR's hardware design forms a tight vertical integration with Intel's underlying software optimization. This ecosystem synergy not only improves the system's overall energy efficiency ratio but also provides a highly attractive alternative for enterprises looking to break free from single-vendor lock-in and build a diversified computing power supply chain, driving the evolution of the AI computing power supply chain towards diversification.

【Developer and Industrial Deployment Implications】

For developers, the engineering integration complexity and the maturity of the software toolchain of GUNNIR's Arc Pro B70 workstation directly determine its deployment value. Using Intel's OpenVINO toolkit, developers can convert PyTorch models into Intermediate Representations (IR) and perform INT8 quantization and hardware-aware optimization on Arc graphics cards, thereby significantly boosting inference speed. However, migration costs still exist, with the main pain point being that custom layers relying on CUDA-specific operators need to be rewritten and adapted using SYCL or OpenCL. Nevertheless, as mainstream model repositories like Hugging Face continuously improve their support for the Intel XPU backend, the threshold for local deployment of mainstream large models is rapidly lowering.

In commercial deployment scenarios, large memory is a core rigid demand for running large models locally. The dual-card Arc Pro B70 workstation can provide a sufficient total VRAM capacity, enabling local quantized inference of 70B-parameter-level open-source large models at INT4 or INT8 precision, without expensive cloud API call fees. This high-density computing power equipment is extremely suitable for vertical industries with strict data privacy requirements. For instance, medical institutions performing localized medical imaging assisted diagnosis and medical record privacy analysis, financial institutions conducting local real-time inference of high-frequency trading data, and design studios performing localized AIGC image and 3D asset generation can all achieve low-latency, high-security industrial deployment.

【Comprehensive Review and Key Takeaways】

The launch of GUNNIR's Arc Pro B70 dual-card liquid-cooled workstation provides a core insight: AI computing power hardware is shifting from mere parameter stacking to scenario-based and form-factor-based refined design. The application of liquid cooling and vapor chamber technology in compact workstations indicates that future high-density AI computing devices will pay more attention to the balance between energy efficiency and space utilization. Hardware vendors are no longer just providing graphics cards, but system-level AI computing solutions that include heat dissipation, power supply, and motherboard layout. This shift will greatly lower the physical barrier to local AI deployment, driving AI computing power from server rooms to desktops, and even to more compact edge computing nodes.

Looking ahead over the next 1 to 2 years, with the continuous iteration of Intel's Xe architecture and the accelerated maturation of software ecosystems like oneAPI and OpenVINO, local AI workstations based on an open ecosystem are expected to capture a significant market share among small and medium-sized enterprises and research institutions. This will not only weaken the moat of proprietary software ecosystems but also promote the localized deployment process of edge-side AI Agents and embodied intelligence control hubs. As the computing power supply chain diversifies and hardware form factors continue to innovate, the democratization process of AI technology will significantly accelerate, providing a more solid and economic underlying computing power support for the intelligent upgrade of various industries.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.