ASUS is making waves in the AI tech world with its newly unveiled mini supercomputer, the Ascent GX10. Designed to provide a remarkable balance of power and affordability, this compact powerhouse can achieve up to 1000 AI TOPs of FP4 precision. The standout feature of this system is the integration of NVIDIA’s cost-effective yet robust GB10 platform.
While the tech community buzzes around NVIDIA’s higher-end GB300 platform, ASUS has strategically opted for the more accessible GB10 to cater to developers eager to explore AI without breaking the bank. This move aligns perfectly with ASUS’s vision of democratizing AI development, making it possible for more innovators to bring their ideas to life.
Under the hood, the Ascent GX10 is built upon the GB10 Grace Blackwell Superchip, featuring a mighty 20-core Arm CPU. This combination ensures that the supercomputer can handle significant workloads, albeit not at the same scale as the GB300, which boasts 20 times the processing power. The real charm of the Ascent GX10 lies in its ability to offer a viable solution for enterprises and individuals seeking cost-effective AI training capabilities.
Equipped with an impressive 128 GB of unified system memory, the Ascent GX10 can tackle AI models with up to 200 billion parameters. This robustness is enhanced by its use of the fifth-generation Tensor cores and FP4 support, which facilitate swifter computations while managing memory efficiently. NVIDIA’s Blackwell GPUs, known for optimal performance with lower memory usage, add to the system’s prowess.
According to KuoWei Chao, the General Manager of ASUS IoT and NUC Business Group, the Ascent GX10 is set to revolutionize industries by bringing AI’s transformative power to the fingertips of developers worldwide. The compact design and powerful capabilities ensure that innovation and the expansion of AI boundaries can happen effortlessly, right from one’s desktop.
The Ascent GX10 also stands out with its advanced connectivity options. It features NVIDIA NVLink-C2C integration, delivering a seamless CPU-GPU memory model with a bandwidth five times that of PCI-E 5.0. Additionally, the option to interconnect two GX10 supercomputers using NVIDIA ConnectX-7 network interface cards allows for handling models with up to twice the parameters, such as Meta Llama 3.1.
While ASUS has yet to reveal the pricing details, anticipation is high for when pre-orders open in the second quarter of 2025. The Ascent GX10 promises to be a game-changer for AI developers, providing an enticing mix of performance, affordability, and accessibility.
