NVIDIA DGX Spark Scales to 8-Node AI Clusters

NVIDIA used GTC Berlin to confirm that DGX Spark, its desk-sized AI box introduced earlier this year, now officially scales to eight-node clusters. That pushes unified memory from the single unit's 128 GB up to a full terabyte across eight machines, networked together over ConnectX-7 RoCE. Four nodes already got you to 512 GB, enough for plenty of work, but a terabyte changes what's actually possible on hardware sitting next to a monitor instead of locked in a data center rack.

The timing tracks with how people are actually using these models now. Developers aren't running one model and waiting for an answer anymore. They're chaining inference calls across several models at once, running multi-agent setups, and testing architectures that need hundreds of gigabytes just to load.

NVIDIA Sync, the company's clustering tool, is meant to make linking nodes less of a chore. Its Cluster Assistant manages setup and device coordination automatically, and support for eight-node configurations in Sync is slated to arrive by the end of the year.

That raises an obvious question: what actually runs on this thing, and how fast.