NVIDIA has officially unveiled the highly anticipated DGX Spark, a compact device that marks the “world’s smallest AI supercomputer.”
Available for $3,999, the powerhouse is specially built to provide developers, researchers, and students with the convenience of accessing data-centre-level computing to build and run advanced AI models locally.
Weighing only 2.6 pounds, the device will be launched on Wednesday, October 15, and will be based on the new GB10 Grace Blackwell Superchip made by NVIDIA.
It is a 20 core Arm-based Grace CPU and a Blackwell GPU, which provides an impressive 1,000 trillion operations per second (1,000 TOPS) of AI compute power.
The performance of this one is optimised for memory-intensive applications such as fine-tuning of large language models, robotics simulation, and generative AI.
The system is geared towards a smooth local development experience.
It includes 128GB shared LPDDR5x RAM, 4TB NVMe storage, and a custom Ubuntu-based DGX OS already loaded with Nvidia's AI software stack.
The embedded NVLink-C2C interconnect offers bandwidth that is five times higher than that of PCIe Gen 5 to make data flow between the CPU and the GPU smooth.
According to the founder of NVIDIA, Jensen Huang, "the new device was connected with the original mission of the company, which was to bring an AI computer to the hands of every developer," he claimed.
"With DGX Spark, we go back to that mission: to bring an AI computer to the hands of every developer to spark the next wave of breakthroughs," he added.
The introduction marks the beginning of the strategic drive to make supercomputer-scaling AI power more affordable, allowing for prototyping and testing without the delays or ongoing expenses of cloud services.
NVIDIA and a few retailers will distribute the DGX Spark.
