New in anti-static bag. In the USA.

Connect two NVIDIA DGX Spark systems with a single cable. No switch. No optics. Plug QSFP-to-QSFP and you have a direct 200GbE RoCE link for multi-node inference and fine-tuning.

This is an Amphenol NJAAKKR-0006: 0.5 m QSFP112 400G-rated passive DAC, 30 AWG. Same family as the Amphenol cables NVIDIA documents for Spark stacking (NJAAKK-N911 / NJAAKK-0006). On a Spark the port tops out at 200 Gb/s, which is what this short copper link is built for.

What you get from two Sparks

• One box: unified 128 GB, models up to ~200B parameters
• Two boxes on this cable: combined memory working set for models up to ~405B (NVIDIA’s dual-Spark figure)
• Point-to-point GPU traffic over ConnectX-7 for distributed inference and fine-tuning
• Sit the two compact units side by side — 0.5 m is the intended length

Works with

• NVIDIA DGX Spark (Founders Edition and standard)
• Other GB10 / Grace Blackwell desktop systems that use the same ConnectX-7 QSFP ports (ASUS, Dell, HP, Lenovo, MSI, and similar OEM boxes)

Specs

• Part: Mellanox NJAAKKR-0006
• Type: Passive DAC (no retimers, no optics — 0 W cable electronics)
• Connectors: QSFP112 to QSFP112
• Length: 0.5 m (1.6 ft)
• Gauge: 30 AWG
• Cable rating: 400G / 112G per lane
• Spark link rate: 200 Gb/s Ethernet + RoCE
• Use: two units = one cable; three-node ring = three cables

Setup notes

Use the same physical QSFP port on both machines. After you plug it in you still assign IPs on the QSFP interfaces and set up SSH — NVIDIA’s “Connect Two Sparks” playbook covers that. This cable does the hardware part.

Not this cable

Not USB-C. Not the 10GbE RJ-45. Not InfiniBand. Spark clustering is the rear ConnectX-7 QSFP ports in Ethernet/RoCE mode.

Ships as a genuine Mellanox NJAAKKR-0006 0.5 m assembly. One cable links two Sparks.