NVIDIA Project DIGITS Renamed DGX Spark, DGX Station Unveiled For AI Domination

NVIDIA DGX Spark and DGX Station on a black background.

The NVIDIA GTC AI Development conference started yesterday, but the big announcements are coming in fast and furious today. This one is all about some powerful AI development hardware, known as DGX Spark and DGX Station. These names come out of left field, but the technology upon which they are built is a known commodity: the Grace Blackwell AI supercomputer architecture, which has expanded into multiple chips. 

Back at CES, NVIDIA announced Project DIGITS, its AI acceleration developer workstation fueled by Grace Blackwell. That’s a combination of its Arm CPU design coupled with a Blackwell GPU, all backed by a whopping 128GB of memory. It’s powerful, small, and laser focused on giving AI developers a suitable workstation to do use for their efforts. Now, Project DIGITS has a new name: DGX Spark. That system is accompanied by a new behemoth based on Blackwell Ultra, DGX Station. 

Project DIGITS Becomes DGX Spark

NVIDIA DGX Spark next to a laptop.
Let’s start briefly with the system we already know. The tiny cube-like DGX Spark computer is built upon the Grace Blackwell-based GB10 chip, which delivers up to 1,000 trillions of operations per second (TOPS) of AI-driving compute power. Rather than have discrete GPU and CPU resources, NVIDIA’s on-package NVLink C2C interconnect provides five times the bandwidth of a PCIe x16 slot. This allows the CPU and GPU to share resources in a single memory pool, not unlike other beefy designs like AMD’s Ryzen 9 MAX+ 395 found in the ASUS ROG Flow z13.

NVIDIA says that DGX Spark is the world’s smallest AI supercomputer. Its focus it to allow data scientists, AI researchers, robotics developers, and students all the power they need to expand their fields of work and study and grow the realm of physical AI. NVIDIA’s Cosmos real-world foundation model and the GR00T robotics foundation model are at the heart of robotics and autonomous vehicles, which is where the company foresees the largest growth over the next several years. The idea is that development would start on these powerful DGX Spark systems and seamlessly transition to the cloud, where the same frameworks are powered by the company’s enormous H100 GPUs. 

DGX Station Scales AI in the Enterprise

NVIDIA DGX Station on a black background.
The DGX Station, on the other hand, is a bit of a surprise, as the company hadn’t previously announced anything quite like it. This behemoth is built on the GB300 Grace Blackwell Ultra Superchip, which takes the concept behind DGX Spark and cranks it up to eleven with 20 petaflops of AI horsepower — that’s 20,000 TOPS. NVIDIA has endowed DGX Station with a whopping 768 GB of memory which should be plenty to hold any model for the foreseeable future.

On top of that, NVIDIA has also given each DGX Station its ConnectX-8 SuperNIC, which makes these systems scale with 800 Gigabits per second of network bandwidth. This allows multiple DGX Stations to be networked together and distribute the load at very high-speeds. These systems were custom-built for NVIDIA’s NIM inference microsoervices and its AI Enterprise platform, which makes distributing those microservices easier at scale. 

DGX Spark and DGX Station Availability

As previously announced when it was still called Project DIGITS, NVIDIA’s DGX Spark will be available in May with a starting price of $3,000. What’s new is that starting today, interested buyers can head on over to NVIDIA.com’s Marketplace and start reserving those systems.

Those interested in DGX Station will have to wait a bit, as systems built on this technology will be available from NVIDIA’s manufacturing partners later this year. Companies that have signed on to build DGX Station products include (alphabetically) ASUS, BOXX, Dell, HP, Lambda, and Supermicro.