ASUS recently unveiled a mini AI supercomputer called the Ascent GX10, powered by the NVIDIA GB10 Grace Blackwell superchip, a powerful and "affordable" desktop computing solution. This device delivers up to 1,000 TOPS (trillions of operations per second) of AI computing power in a compact form, designed to meet the growing demand for AI model development while remaining both affordable and portable.

At the core of the Ascent GX10 is the NVIDIA GB10 chip, featuring a 20-core Arm-architecture Grace CPU and a Blackwell architecture-based GPU. These components share unified memory access through NVLink-C2C high-speed interconnect technology, offering five times the bandwidth of PCIe 5.0. While NVIDIA's high-end GB300 platform is designed for hyperscale AI tasks, delivering up to 20,000 TOPS, the GB10 emphasizes a balance between price and performance, suitable for most small to medium-sized AI workloads. The Ascent GX10 is equipped with 128 GB of LPDDR5X unified memory, supporting developers in training, testing, and inference of AI models with up to 200 billion parameters, such as open-source models like DeepSeek, Grok, or Llama 2.

Regarding hardware design, the Ascent GX10 maximizes the potential of Blackwell GPUs' fifth-generation Tensor cores and FP4 precision support. The FP4 format enhances computing speed and minimizes memory usage while maintaining precision, which is essential for real-time inference and model fine-tuning. Additionally, the 4TB built-in NVMe storage is ample for large datasets and model files. ASUS has integrated the NVIDIA ConnectX-7 network interface, allowing users to expand compute power through dual interconnections, managing models like Meta's Llama 3.1, which has 405 billion parameters.

The mini supercomputer is also notable for its exterior design. Featuring a sleek white body with a power button on the front and a unique top design, it's both functional and visually appealing. Specific interface details remain unannounced; however, based on NVIDIA's Project DIGITS, one can expect common connectivity options such as USB4 Type-C, Wi-Fi, Ethernet, and HDMI 2.1, facilitating easy setup for monitors, keyboards, and mice.

Weichao Guo, General Manager of ASUS IoT and NUC Business Unit, remarked, "AI is reshaping every industry, and the Ascent GX10 is designed to put this transformative power into developers' hands. By harnessing the NVIDIA Grace Blackwell chip, we've created a compact yet powerful tool that enables data scientists and AI researchers to innovate within a desktop environment." This highlights the Ascent GX10's target audience—individuals or small teams requiring high-performance computing resources locally, but constrained by budget.

Meanwhile, competition in the AI supercomputer space is intensifying. South Korea's plan to launch a national-level AI supercomputer has been postponed, while NVIDIA's Project DIGITS, introduced at CES 2025, also utilizes the GB10 chip, priced at about $3,000. In contrast, the Ascent GX10 is expected to be available for pre-orders in Q2 2025, with pricing anticipated to be competitive. Analysts suggest that products like the Ascent GX10 will bridge the gap between cloud services and costly data center solutions as generative AI models grow more complex, increasing the demand for local compute and memory capabilities.

Technically, the GB10 chip's Grace CPU is built on the Arm Neoverse V2 architecture, with 10 Cortex-X925 high-performance cores and 10 Cortex-A725 efficiency cores, offering multitasking and energy efficiency. The Blackwell GPU supports CUDA cores and light-tracing technology, achieving 1 PFLOPS FP4 arithmetic, perfect for AI inference and lightweight training tasks. Notably, the Ascent GX10's energy efficiency requires only standard power connections, a substantial improvement over traditional servers' requirements.

The introduction of the Ascent GX10 is timely. By early 2025, the AI hardware market is experiencing a new surge, with AMD and Intel also entering the AI workstation market, while NVIDIA maintains its GPU ecosystem leadership. ASUS's collaboration with NVIDIA to integrate the GB10 chip into desktop products not only expands its AIoT portfolio but also offers developers a flexible platform for experimentation. Whether for prototyping, model tuning, or running inference tasks directly, the Ascent GX10 delivers exceptional performance in a small package.

For technology enthusiasts, the Ascent GX10 signifies more than just hardware—it's a glimpse into the trend of AI technology transitioning from data centers to desktops. More details, including full specs and release plans, are expected to be announced soon, making it worthwhile to stay informed.