Saying goodbye to conventional laptops Everyone is taken aback by ASUS's futuristic mini-supercomputer
Today, ASUS announced the ASUS Ascent GX10, its newest small supercomputer, which has the potential to fundamentally alter the field of local AI development. With the help of the potent NVIDIA GB10 Grace Blackwell processor, this device, which is intended for developers, AI researchers, and data scientists, brings datacenter computing power to a desktop form factor.
Conventional systems are starting to show their limitations as generative AI models continue to increase in complexity and scale. Previously limited to specialized centers or cloud environments, the hardware infrastructure needed to create, tune, and run models with hundreds of billions of parameters is becoming more and more demanding. This paradigm is intended to be altered by ASUS's proposal.
Thanks to its 128GB of unified system memory, the new Ascent GX10 can achieve up to 1,000 TOPS of AI performance. Compared to what desktops have so far provided, this capability makes it possible to manage AI models with up to 200 billion parameters directly from the desktop.
The potent NVIDIA GB10 chip, a CPU/GPU combination that makes use of the Grace Blackwell architecture and is especially made to optimize performance for AI tasks in confined spaces, is at the center of this device. The 20-core Grace Arm CPU speeds up operations like data preprocessing and model formatting, while the integrated Blackwell GPU incorporates 5th-generation Tensor cores and FP4 support. With transfer speeds up to five times faster than PCIe 5.0, the system achieves effective CPU/GPU integration thanks to the NVLink-C2C interconnect.
In addition to allowing native work with models with up to 70 billion parameters, this capability allows performance to be scaled to handle even more ambitious models, like the recently released Llama 3.1 with 405 billion parameters, by connecting two Ascent GX10 modules using integrated NVIDIA ConnectX network cards.
In addition to its technical features, the device's methodology aims to increase the productivity of the developer workflow.
For experimentation, Ascent GX10 is a particularly strong and affordable on-premises platform that frees up more costly and sophisticated computing resources for clusters used for training or deployment. Furthermore, models created on GX10 can be easily transferred to cloud infrastructures like NVIDIA DGX Cloud thanks to support for the NVIDIA software ecosystem, all without requiring major code changes.
This proposal is more than just a technical development in ASUS's eyes. In the words of ASUS IoT and NUC Group general manager Cui Zhao:
All industries are undergoing a revolution thanks to artificial intelligence, and the ASUS Ascent GX10 is made to make this revolutionary power accessible to all developers. Through the integration of the potent NVIDIA Grace Blackwell chip, we provide a potent, small tool that facilitates creativity directly from the desktop.