Use NVIDIA Sync clustering so larger models and training can span nodes.
Some workloads need more compute and memory than a single DGX Spark can provide, but you can still run those workloads if you "cluster" two or more Sparks. Clustering devices is the process of physically connecting them with high-speed cables (and potentially a switch) and then configuring a particular network across those cables. You can't really do it over a wireless connection because the data will not move between devices quickly enough.
The primary component enabling a high-speed cluster across Sparks is the ConnectX-7 network adapter that each Spark comes with. The ConnectX-7 adapter, often called a "NIC", is a software and hardware layer that enables data transfer directly between memory across devices through RDMA.
If you physically cable two or more Sparks with QSFP cables and properly configure the ConnectX-7 network, you can have high-speed super computing cluster on your desktop.
This playbook shows you two paths to configure the ConnectX-7 network across two to four DGX Spark devices.
Both paths start with following the Cable the Devices tab to make sure the devices are appropriately connected.
json files. Advancedbash scripts. AdvancedCheck the table below to see if this playbook is for your hardware.
| Hardware platform | OS | Memory | Recommended default local settings | Multi-node capable hardware |
|---|---|---|---|---|
| DGX Spark | DGX OS (Linux) | 64 GB or 128 GB Unified Memory | Direct or switch QSFP links | ✅ (200GbE QSFP) |
Hardware requirements
Software requirements
sudo privileges on each device.These files are only needed for the manual path. You can find them in this playbook's assets folder.
spark_cluster_setup — network and SSH setup for the manual path