NVIDIA's Palm-Sized AI Supercomputer Goes on Sale, with Jensen Huang Personally Delivering to Musk!

Jensen Huang personally delivered the first batch of DGX Spark to SpaceX Chief Engineer Musk. This palm-sized supercomputer is ushering in a new era of AI popularization.

"From a single spark, ignite the world of intelligence!" At the delivery ceremony for NVIDIA's DGX Spark, SpaceX Chief Engineer Elon Musk wrote these words on a gift presented to Jensen Huang.

On October 15th, NVIDIA officially delivered the world's smallest AI supercomputer to date, the DGX Spark. This palm-sized device brings data-center-level computing power to every developer's desktop.

01 A Symbolic Delivery

Jensen Huang traveled to SpaceX's Starship base in Texas to personally deliver the first batch of DGX Spark units to Musk. This scene felt like a historical echo.

In 2016, Huang delivered the first DGX-1 AI supercomputer to OpenAI, which was still a small startup at the time.

Nearly a decade later, he once again placed cutting-edge AI computing equipment into the hands of the same visionary.

Musk's return gift was inscribed: "From a single spark, ignite the world of intelligence! To Jensen Huang, to the stars!"

The phrase "Ad astra" (Latin for "to the stars") aptly expressed the shared desire of these two tech pioneers to explore the unknown.

02 A Supercomputer in the Palm of Your Hand

The DGX Spark weighs only 1.18 kilograms, with dimensions similar to a Mac mini. Yet within this champagne gold metal shell lies astonishing computing capability.

It is equipped with NVIDIA's GB10 Grace Blackwell superchip, delivering 1 PFLOPS (one petaflop, or one thousand trillion floating-point operations per second) of AI computing power in sparsity FP4 precision.

This means it packs the computing power that once required clusters of machines into a desktop-level device.

The 128GB of coherent unified memory is a major highlight of the DGX Spark. The CPU and GPU share the same physical memory space, seamlessly accessing each other's data.

Developers can load entire models with hundreds of billions of parameters into this 128GB of memory at once for operation, eliminating the need to copy data between system memory and VRAM.

03 Impressive Performance

The performance of the DGX Spark is remarkable. According to evaluation data, when running the GPT-OSS 20B model, the Spark's prefill throughput reaches approximately 2053 tokens/second.

For smaller models, like the Llama 3.1 8B model, the Spark's prefill speed reaches about 7991 tokens/second under a single request.

Even more surprising is that, via the dual QSFP network interfaces (200Gb/s bandwidth) on the back, two Spark units can be directly connected to form a two-node cluster.

NVIDIA states that a dual-linked Spark setup can handle models with up to 405B parameters (FP4 precision), approaching the scale of the largest currently available open-source large models.

04 Designed for Developers

The DGX Spark is not designed for ordinary consumers, but is a professional tool for AI developers, researchers, and data scientists.

Jensen Huang stated, "In 2016, we developed the DGX-1 to provide AI researchers with their own dedicated supercomputer. Today, the DGX Spark allows us to return to that original intention – placing an AI computer in the hands of every developer, to spark the next wave of breakthroughs."

The DGX Spark runs NVIDIA's DGX OS, a customized version of Ubuntu Linux, pre-configured with a complete AI software stack.

It is ready to use out-of-the-box, providing immediate access to the NVIDIA AI ecosystem of tools and resources, including models, libraries, and NVIDIA NIM microservices.

05 The Supply Chain Behind the Scenes

Taiwan tech companies played a key role in the manufacturing process of the DGX Spark.

MediaTek was responsible for the GB10 superchip, optimizing it for NVIDIA in terms of CPU connectivity, ISP, and power management.

This chip is built using TSMC's 4-nanometer process.

Brands like Acer, ASUS, Gigabyte, and MSI will also launch DGX Spark series systems, leveraging their respective strengths in mechanical design, motherboards, and cooling technology to target the AI developer and small-to-medium enterprise market.

06 The Future of Computing Power

The launch of the DGX Spark marks the beginning of the "localization" trend for AI computing power.

It will extend AI computing beyond cloud data centers to the desktops of enterprises and developers.

As AI models evolve from text and vision to reasoning and autonomy, the growth rate of computational demands is surpassing the expansion speed of cloud computing capabilities.

NVIDIA's bet is that bringing supercomputing capability to the user's vicinity is crucial for maintaining this pace of development.

The DGX Spark is precisely like a single spark. Priced at just $3,999 (approximately ¥28,500 RMB), it possesses computing power that once existed only within data centers.

From the miniature device in Jensen Huang's hand to the desks of developers worldwide, a new era of AI popularization is beginning.

As Musk said, "From a single spark, ignite the world of intelligence."