This computer is smaller than a router but still crushes the competition

NVIDIA launches a compact AI computer that outperforms traditional PCs in specific tasks – and it takes up less space than a router.

NVIDIA has unveiled DGX Spark, a compact AI computer designed for local AI computing.

Measuring just 150 × 150 × 50.5 mm and weighing around 1.2 kg, it features hardware previously only found in servers.

It is aimed at developers and researchers who want to work with large language models without relying on cloud services.

The computer is built around the Grace Blackwell GB10 superchip.

It combines a 20-core ARM processor and an AI accelerator with fifth generation Tensor cores and fourth generation RT cores.

The big technical advantage is the shared memory of 128 GB LPDDR5X, which both CPU and GPU share.

This eliminates the need to copy data between RAM and VRAM and enables faster and more efficient processing of large models.

Ready to use with full software suite

DGX Spark comes with DGX OS, a Linux-based system, and the full NVIDIA AI software stack installed out of the box.

This means users can quickly get started training, tuning and running models without having to manually configure the environment first.

It supports both large language models and multimodal models that work with text and images simultaneously.

Connectivity options include HDMI, 10 GbE LAN and four USB-C ports.

It supports up to 4 TB of NVMe storage and can be paired with another DGX Spark to form a small AI cluster.

This makes it possible to work with even larger models or increase efficiency.

Unlike traditional PCs with separate CPU and graphics cards, DGX Spark has a unified architecture.

This provides better performance for AI purposes, especially for inference and light training.

At the same time, power consumption is limited to 240 watts, making it more energy efficient than a powerful desktop PC.

https://twitter.com/_ksasaki/status/1990008695636578471

However, some users have reported heat issues and lower performance than advertised.

Still, DGX Spark marks a new approach to local AI processing and focuses on efficient performance in a very compact format.