Is this page helpful?

Spark Stacking#

Connecting DGX Spark Systems into a Cluster#

Overview#

This guide explains how to connect multiple DGX Spark systems into a compute cluster using simplified networking configuration and a QSFP/CX7 cable for high-performance interconnect.

The goal is to enable distributed workloads across Grace Blackwell GPUs using MPI (for inter-process CPU communication) and NCCL v2.28.3 (for GPU-accelerated collective operations).

Connect the QSFP/CX7 Cable#

Before configuring networking, connect the two DGX Spark systems with the approved QSFP/CX7 cable. Each unit has two ConnectX-7 ports on the rear panel; see Connectivity and I/O for port locations.

  1. On each DGX Spark, locate the two QSFP/CX7 ports on the rear panel.

  2. Plug one end of the cable into a ConnectX-7 port on the first unit, and plug the other end into the same ConnectX-7 port on the second unit (for example, both left ports or both right ports when viewed from the rear of the unit).

  3. Orient the connector so the pull-tab (ring tab) faces the top of the DGX Spark (toward the top surface of the unit), as shown in the following figure.

    DGX Spark rear panel with QSFP/CX7 cable inserted; pull-tab faces the top of the unit

  4. Align the connector with the port and insert it until it seats fully. When aligned correctly, the connector slides in without force.

  5. To remove the cable, pull the ring tab straight out. Removal should be smooth when the connector is fully seated.

Warning

Do not force the QSFP/CX7 connector. If it does not slide in smoothly, stop, verify tab orientation and port alignment, and try again. Forcing an upside-down or misaligned connector can damage the port.

Connect Multiple DGX Spark Systems#

Refer to the following playbooks for information on various connection options:

Next Steps#

Once tested, this configuration can be scaled to support:

  • Job orchestration with Slurm or Kubernetes

  • Containerized execution with Singularity or Docker