The days of High-Performance Computing demanding a server rack or full-tower workstation are gone. 

Introducing the NVIDIA DGX Spark — a lunchbox-sized AI supercomputer capable of 1 petaFLOP of throughput for the price of a high-end gaming PC. 

When NVIDIA introduced the DGX Spark platform, it quickly became the most recognizable compact AI system on the market. It is small, somewhat funny-looking, and supports daisy-chaining multiple units to build a local dual-node cluster. It also comes preinstalled with the same proprietary NVIDIA software stack that is used with their flagship DGX servers.

At Central Computers, you can find the reference model DGX Sparks, in addition to other variations of the platform from companies such as ASUS, Gigabyte, and MSI, starting at $2,999.99 for 1TB storage models up to $3,999.99 for 4TB. 

What is the DGX Spark made for, exactly?

The DGX Spark is NVIDIA’s reference design for a 1 petaFLOP-capable desktop AI workstation in a mini-PC form factor. Within just one liter of volume, it houses the GB10 Grace Blackwell superchip alongside 128GB of unified memory, up to 4TB of NVMe storage, and high-speed networking via ConnectX-7. It’s purpose-built for developing and running local machine learning models as well as fine-tuning data-science workflows. 

Because it shares the same NVIDIA AI software environment as full-size DGX servers, the DGX Spark is an ideal platform for developers looking to prototype and validate workloads locally before scaling them into datacenter deployments.

Why are there so many different versions?

Since NVIDIA opened the DGX Spark platform to OEM partners, several manufacturers now produce their own GB-10-based systems, running the same hardware and operating system, same memory and storage configuration options, but with different chassis designs among some other smaller details. We will break down each of the most important differences in this article, so that you have an easier time choosing your DGX Spark system when shopping at Central Computers.

What makes them different?

At a component and software level, all DGX Spark partner systems are built on the same NVIDIA GB10 Grace Blackwell superchip platform, offer the same connectivity and I/O, use the same DGX Spark OS, and cost roughly the same as their alternatives. Despite having differing chassis designs, they even share nearly identical dimensions. 

In practice, this means that model selection mostly comes down to:

1. Storage configuration
   
   The reference model DGX Spark products (NVIDIA + PNY) do not offer storage options outside of the default 4TB NVMe drive. If you do not have a particular need for a 4TB drive and wish to save money by going with a cheaper drive, you may want to consider the ASUS Ascent GX10, which we offer in 1TB, 2TB, and 4TB configurations. However, it’s important to note that the 1TB and 2TB versions of the ASUS Ascent GX10 only support PCIe Gen 4.0, as opposed to the 4TB version which does offer PCIe Gen 5.0. The rest of the models being compared here all use PCIe Gen 5.0 4TB SSDs.
2. Airflow setup and visual design
   
   

   Each DGX Spark system we carry at Central Computers shares a similar integrated cooling system, with minimal differences in thermal performance. Anybody who is looking to use a custom dual-system chassis should check the compatibility of their chassis to make sure the model they’re purchasing will fit, as the dimensions and screw placements can vary between models. 

   The reference model DGX Sparks use the famous NVIDIA DGX workstation metal foam design for the front intake grille, and sport a unique gold finish. Other systems from ASUS, Gigabyte, and MSI are much more reminiscent of a typical mini PC — albeit with distinct styling elements unique to each brand.

3. Buyer's preference of vendor
   
   

   Everybody has their own preferences for vendors; whether you’re a user who got burnt by one manufacturer’s customer service, or an enterprise buyer who wants to maintain homogeneity in their workstation fleet, it’s important to do your own research and consider past experiences to determine which model you want to buy.

   For considering users: All DGX Spark systems currently carried by Central Computers include a standard 1-year limited manufacturer warranty.

   NVIDIA and PNY DGX Spark units are processed through NVIDIA, while ASUS, Gigabyte, and MSI systems are serviced directly through their respective manufacturer channels.

Scaling with dual-node clusters

While a single DGX Spark system is powerful on its own, two units can be interconnected to create a dual-node cluster thanks to the ConnectX-7 port present on the I/O. By linking them with a high-speed 200G Mellanox QSFP cable, distributed workloads can be run between the two systems, effectively doubling available compute resources and memory capacity. 

Pictured below is an example of a dual-node ASUS Ascent GX10 system as part of a project that our Corporate IT Service department has been working on.

Where does the DGX Spark fit into my workflow?

The DGX Spark is best understood as a bridge between midrange homelab workstations and full-on datacenter servers. Rather than replacing either outright, it’s a great platform for testing your memory-intensive NVIDIA AI pipeline in an office, lab, or any research environment where you might not be able to haul a full-size workstation and/or have power limitations. 

Central Computers carries a large lineup of NVIDIA DGX Spark and other GB10-based mini PC systems from companies such as ASUS, Gigabyte, and MSI. 

You can browse available models, storage configurations, and even compatible QSFP networking cables at our rapidly growing AI Supercomputer catalog. 

For businesses wanting to add DGX Spark systems to their pipeline, our Corporate Sales team is always available to assist with consultations and procurement for any projects you might have.