
ASUS has introduced a groundbreaking compact AI supercomputer capable of achieving up to 1000 AI TOPs of FP4 precision, utilizing the high-performance GB10 from NVIDIA.
Introducing the ASUS Ascent GX10 Mini Supercomputer: A Compact Powerhouse for AI Training
While the spotlight is often on NVIDIA’s advanced GB300 platform, the more economical and efficient GB10 has caught the attention of ASUS. This has led to the creation of the “Ascent GX10, ”a mini supercomputer designed to offer robust capabilities for training and executing demanding AI tasks at a fraction of the cost. During the GTC event, among several announcements from leading hardware manufacturers, ASUS unveiled the Ascent GX10, featuring NVIDIA’s GB10 Grace Blackwell Superchip paired with a powerful 20-core Arm CPU.
The impressive CPU-GPU combination equips the Ascent GX10 to deliver up to 1, 000 AI TOPs (Tera Operations Per Second) of performance, positioning it well for handling extensive AI workloads. In contrast, the GB300 excels in extreme processing needs with a staggering 20, 000 AI TOPs, targeting more intensive applications. The Ascent GX10, however, aims to provide a value-driven solution for developers eager to train AI models without breaking the bank.

Equipped with 128 GB of coherent unified system memory, the Ascent GX10 is tailored for developing, experimenting with, and inference of AI models containing up to 200 billion parameters. Built upon the NVIDIA GB10 architecture, it harnesses fifth-generation Tensor cores and FP4 support, which enhances computation speed while ensuring efficient memory usage. This innovative design allows NVIDIA’s Blackwell GPUs to excel particularly in scenarios demanding lower memory footprints.

The intended focus of the ASUS Ascent GX10 is to provide a cost-effective solution for AI development. According to KuoWei Chao, General Manager of ASUS IoT and NUC Business Group:
AI is transforming every industry, and the ASUS Ascent GX10 is designed to bring this transformative power to every developer’s fingertips.
By integrating the NVIDIA Grace Blackwell Superchip, we are providing a powerful yet compact tool that enables developers, data scientists, and AI researchers to innovate and push the boundaries of AI right from their desks.
Moreover, the Ascent GX10 facilitates high-speed connectivity through NVIDIA NVLink-C2C, which establishes a cohesive CPU-GPU memory model capable of offering five times the bandwidth of PCI-E 5.0. Developers can connect two GX10 supercomputers via the NVIDIA ConnectX-7 network interface cards, enabling them to manage models with parameters reaching approximately 405 billion, comparable to the Meta Llama 3.1.
While ASUS has yet to reveal pricing, pre-orders for the Ascent GX10 are expected to open in the second quarter of 2025.
Leave a Reply ▼