This is where most people get stuck. NVIDIA's own documentation references specific Amphenol and Luxshare part numbers, but those aren't always easy to source. The compatible cable you need is:
Part number: Q56-200G-CU0-5 A 0.5-meter, 200G QSFP56 Passive Direct Attach Copper (DAC) Twinax Cable
Why 0.5 meters? Because you'll be placing the two DGX Sparks right next to each other — they're compact desktop units. A longer cable adds unnecessary signal degradation. The 0.5m length is optimized for exactly this use case.
You plug one end into the ConnectX-7 QSFP port on the back of one DGX Spark and the other end into the same port on the second unit. No switch required. No SFP adapters. Just direct copper.
We stock the Q56-200G-CU0-5 and ship fast: Shop it here →
What Software Do You Need?
Once the cable is in, you'll configure the ConnectX-7 ports as a direct Ethernet link between the two machines. NVIDIA's official Spark Stacking guide walks through networking configuration using MPI for inter-process CPU communication and NCCL for GPU-accelerated collective operations. NVIDIA From there you can use vLLM, NIM, or other distributed inference frameworks.
Important Limitations to Know
NVIDIA currently supports stacking of up to 2 DGX Sparks with an approved QSFP cable. nvidia You can't chain three or more units in series using this method. Also note that the DGX Spark's ConnectX-7 ports operate in Ethernet mode only — not InfiniBand.
Quick Checklist
- ✅ Two DGX Spark units (Founders Edition or standard)
- ✅ One Q56-200G-CU0-5 DAC cable (0.5m, 200G QSFP56)
- ✅ Both units running DGX OS / Ubuntu 24.04 or later
- ✅ NCCL v2.28.3+ and your inference framework of choice
That's it. No switches, no transceivers, no extra hardware.