ASUS

Ascent GX10: the "mini" personal supercomputer for creating AI applications

Created to help programmers, researchers, and data scientists work with artificial intelligence, Asus has launched the Ascent GX10 personal supercomputer in Brazil. The model is powered by Nvidia's GB10 Grace Blackwell superchip, offering 1 petaflop of AI performance.

The model features a scalable architecture and includes a system for fast communication between the CPU and GPU memory, via Nvidia's NVLink-C2C. The computer is also designed for maximum thermal efficiency, nearly twice that of similar compact systems.

Almost simultaneously with the Dell Pro Max GB10a, Asus introduced a new mini PC based on the Nvidia GB10 Grace Blackwell ARM chip. This mini PC features a simple white case with a power button on the front and a distinctive design on the top.

The Nvidia chip features 10 ARM Cortex-X925 performance cores, as well as 10 additional Cortex-A725 cores. The processor is connected via NVLink interconnect to a graphics chip based on the current Nvidia Blackwell architecture. This allows both the processor and the graphics chip to access memory, which can be configured with a capacity of up to 128 GB. This system is reportedly capable of handling AI models with up to 200 billion parameters.

The computer dedicated to artificial intelligence tasks aims to meet the demands of creating AI applications. The Nvidia superchip includes a 20-core Grace CPU and a Blackwell GPU, capable of processing 1 petaflop of AI performance for inference and model fine-tuning. It also features 128 GB of unified memory, allowing it to work on models with up to 200 billion parameters directly from the machine.

The Ascent GX10 also demonstrates that supercomputers aren't measured in inches, being compact, at about 15 cm2, allowing it to be placed anywhere on a desktop. The computer aims to be a full-stack AI solution, accelerated by Nvidia's AI software stack. Its main functions include prototyping, fine-tuning, and inference for developing applications in robotics, computer vision, and visual language models.

If you need even more power, the scalable architecture allows you to connect two Ascent GX10 units via high-speed connectX-7 network interfaces, instantly doubling AI performance to 2 petaflops, up to 256 GB of unified memory, and 8 TB of storage. Asus hasn't revealed pricing for the machine, but claims it's a powerful solution for local AI training, with reduced costs, enterprise-grade security, and the ability to keep sensitive data on-premises.

Nvidia advertises AI performance of 1,000 TOPS, which is about a third of the performance of a GeForce RTX 5090. However, the large amount of memory is suitable for larger models that can be worked on without swapping and losing performance along the way.

Additionally, users can combine multiple Asus Ascent GX10 units if they need even more memory. The mini PC is also equipped with up to 4 TB of SSD storage. Asus has not yet confirmed other technical details and specifications, such as connectivity options.

Pricing and Availability...Asus also did not provide any details about the price or release date of the Ascent GX10. Nvidia's technically comparable Mini PC, Project Digits, will launch in May for $2,999, so it's reasonable to assume the Asus system will be similarly priced.

mundophone