USA
Connecting two NVIDIA DGX Spark systems for bigger AI models? Many customers ask for an “NVLink cable”. There isn’t one, and you don’t need one. Each Spark has a built-in ConnectX-7 network controller, so a single validated passive DAC cable links the two units directly, with no switch. This guide covers which cable to buy, the NVIDIA-listed part numbers, the cables to avoid, and how to set it up.
Many customers ask for an “NVLink cable”. There isn’t one, and you don’t need one. Here is the single cable that does the job, the validated part numbers and the mistakes to avoid.
Each DGX Spark has 128 GB of unified memory. Clustering two gives you a larger combined memory pool, so you can load models that won’t fit on a single unit.
It’s a memory-capacity upgrade, not a speed upgrade, because the link between the boxes is slower than the memory inside each one.
Pooling memory across two units lets you load models that are too big for one box. Expect the cross-box link to set the pace, so plan around capacity rather than raw throughput.
Capacity upgrade — not a speed upgrade.
The DGX Spark has no external NVLink connector. Every unit carries an NVIDIA ConnectX-7 network controller with two QSFP112 ports.
There is nothing to bridge on the outside of the box, so an “NVLink cable” doesn’t exist for it.
The built-in ConnectX-7 gives you the ports. You don’t need to add a network card.
Connect the units directly through those ports. No switch is needed.
NVIDIA’s documentation lists validated cables for the Spark’s ports. These are the ones to look for.
The standard Spark stacking cable.
Same cable, slightly longer.
NVIDIA-listed equivalent.
Tip: for two units side by side on a desk, the short passive cable is the right choice. A longer or active cable adds cost and gains you nothing.
Most of the “200G QSFP” cables you’ll find in a catalogue are the wrong type for this job.
MCP1650 seriesMellanox HDR 200G QSFP56. They look right on paper, but they are a different cable class from the Spark’s QSFP112 ports and aren’t on NVIDIA’s validated list.
MMA1Z00-NS400 and similar are for fiber runs between racks. Two Sparks on the same desk don’t need them.
The ConnectX-7 is already built into the Spark. You don’t need to add one.
The same ports support small clusters. Pick a size to see how it connects.
Cables that are only “compatible” are the most common cause of hard-to-diagnose link failures. Confirm yours is one of the validated parts above before changing any configuration.
We supply the NVIDIA DGX Spark Founders Edition and the matching interconnect cable, so your cluster arrives with parts that work together. Tell us how many units you’re clustering.
Customers often ask for an NVLink cable to join two DGX Spark units. The Spark has no external NVLink connector. Its built-in ConnectX-7 network controller links two units directly with one passive DAC cable, and no switch is needed.
Look for the Amphenol NJAAKK-N911 (400 mm, the standard stacking cable), the Amphenol NJAAKK0006 (0.5 m) or the Luxshare LMTQF022-SD-R. Plug it into QSFP Port 0 on each unit. For two units on one desk, the short passive cable is all you need.
Skip the MCP1650 series (HDR 200G QSFP56), which is a different cable class and not on NVIDIA’s validated list. Skip 400G optical transceivers such as the MMA1Z00-NS400, which are for rack-to-rack fiber. Don’t add a network card, because the ConnectX-7 is already built in.
The same ports support small clusters. Three units link directly in a ring with three cables, each box using both ports. Beyond three, you’ll want a switch. NetworkDistri supplies the DGX Spark Founders Edition with the matching cable. Tell us how many units you’re clustering and we’ll quote it.
