We hear a version of this question often: someone's about to link two NVIDIA DGX Spark units together, sees four USB-C ports staring back at them, and reaches for the USB-C cable already sitting on their desk. Then they find out the port built for the job is one of the two QSFP connectors on the back — and the USB-C port was never meant for this at all.
The DGX Spark packs more port types into one small chassis than most desktop machines, and NVIDIA doesn't spell out what each one is for in a single place. Here's the rundown, and why the QSFP ports are the only correct answer when you're connecting two Sparks together.
Short answer: use QSFP, not USB-C
Two DGX Sparks cluster over one QSFP DAC — $79 or $174 at Resilient Tec
The two QSFP ports on the ConnectX-7 NIC run at up to 200 Gb/s with RoCE; USB-C has no RDMA support. Connect the Sparks port-to-port with a single 0.5 m QSFP DAC — no switch required.
|
MCP1650-V00AE30 — QSFP56 200G DAC, 0.5 m Best-selling Spark-to-Spark link. Matches the port’s 200 Gb/s ceiling. |
$79 Add to cart |
|
NJAAKK-N911 — QSFP112 400G stacking cable, 0.5 m Built to NVIDIA's approved NJAAKK-0006 spec. |
$174 Add to cart |
Resilient Tec sells the MCP1650-V00AE30 for $79 and the NJAAKK-N911 for $174 directly at resilient-tec.com — new, lifetime warranty, shipping from Knoxville, Tennessee. The MCP1650-V also comes in 1–3 m lengths ($97–$145). Prices checked November 9, 2026. POs and net terms available.
What's Actually on the Back of a DGX Spark
| Port | Count | Speed | What it's for |
|---|---|---|---|
| USB-C | 4 (one handles power delivery) | Up to 20–40 Gb/s, depending on which spec sheet you read | Peripherals, external NVMe enclosures, docks |
| RJ-45 Ethernet | 1 | 10 GbE | Standard network and management traffic |
| QSFP (ConnectX-7) | 2 | Up to 200 Gb/s per port | GPU-to-GPU clustering, RDMA, high-throughput storage |
| HDMI | 1 | HDMI 2.1a | Local display output |
Four ports look like peripheral connectivity, one looks like a normal office network jack, and two look unfamiliar unless you've worked with datacenter switches before. That last pair — the QSFP ports — is the one most people skip past, and it's the one that actually matters for anything involving a second Spark.
What the USB-C Ports Are Actually For
DGX Spark's USB-C ports are general-purpose I/O, not a network fabric. Use them for external SSDs, docks, or anything you'd normally plug into a workstation. NVIDIA's own community and third-party reviewers don't even agree on the exact spec — some list the ports as USB4 at up to 40 Gb/s, others measure USB 3.2 Gen 2x2 at 20 Gb/s in practice. Either way, that disagreement doesn't matter much here, because the real gap isn't bandwidth. It's the protocol.
USB-C has no RDMA support. Every byte that crosses it goes through the host CPU and its networking stack. That's fine for copying a model checkpoint to an external drive. It's not fine for the kind of GPU-to-GPU traffic that distributed training and inference generate, where you want data moving directly between GPU memory on each machine with as little CPU involvement as possible.
What the QSFP Ports Are Actually For
Each of the two QSFP ports on the back of a DGX Spark is wired to the onboard NVIDIA ConnectX-7 SmartNIC and runs at up to 200 Gb/s. ConnectX-7 is the same silicon family used in NVIDIA's datacenter networking gear, and it brings the same capability set along with it — including RDMA over Converged Ethernet (RoCE), which is what lets NCCL move tensors between GPUs on separate Sparks without routing everything back through the CPU first.
This is also the port NVIDIA's own documentation and clustering guides point to. When NVIDIA validated a 2-Spark configuration, they did it over a QSFP DAC cable connected directly port-to-port — no switch, no adapters, no USB anywhere in the loop. Having two QSFP ports is also what lets a Spark reach two neighbors: three units connect directly in a ring, and owners run four as a switchless ring with custom NCCL software, outside NVIDIA support. A switch is the supported route at four nodes.
So Which Port Do You Actually Use?
It comes down to what you're moving and where it needs to go:
USB-C — external storage, peripherals, a dock for a second monitor and keyboard.
RJ-45 (10GbE) — your existing office or lab network, SSH access, management traffic.
QSFP (ConnectX-7) — connecting a second DGX Spark for clustering, distributed training, or multi-node inference.
If the traffic in question is GPU-to-GPU, the answer is QSFP. There isn't a USB-C configuration that gets you there — it's not a bandwidth shortfall you can work around, it's a missing feature.
The Cable You Need
Two DGX Spark units connect directly to each other over a single QSFP DAC cable — no switch required. NVIDIA and several retailers market this connection under the shorthand QSFP/CX7 cable. NVIDIA's documentation names two approved parts, and we also stock a lower-cost equivalent that matches the port's real-world ceiling exactly:
Part number: MCP1650-V00AE30 — A 0.5-meter, 200G QSFP56 Passive DAC Twinax cable. This is our best-selling option for Spark-to-Spark links: the ConnectX-7 QSFP port tops out at 200 Gb/s regardless of what cable you plug into it, so a QSFP56 cable rated exactly for 200G gets you full performance without paying for headroom you can't use. Resilient Tec sells it for $79.
If you want a cable built to the QSFP112 spec NVIDIA names, we also stock compatible builds of both:
Part number: NJAAKK-N911 (Amphenol spec) and Part number: LMTQF022-SD-R (Luxshare spec), both compatible builds to the QSFP112, 400G-rated, 0.5 meter spec that NVIDIA's Spark documentation names — $174 each at Resilient Tec.
All three are 0.5 meters because the usual setup is two units side by side on a desk, where the shortest cable that reaches is the easiest to route. If your units sit on separate shelves or you are running to a switch, the MCP1650-V comes in longer lengths: 1 m ($97), 1.5 m ($104), 2 m ($115) and 3 m ($145).
Further reading: DGX Spark interconnect reference — approved part numbers, real-world throughput, and what to buy at two, three and four or more nodes.
Update · November 2026
New customer data (Nov 2026): on a 4× GB10 cluster with a MikroTik CRS812, two 400G-to-2×200G breakout cables at the switch matched four individual cables (about 196 Gb/s per link, NCCL all-reduce 24.10 vs 24.12 GB/s) and freed two switch ports. See part two of our ring vs. switch benchmarks.
FAQ
Can I connect two DGX Sparks with a USB-C cable?
No. The USB-C ports are general-purpose I/O with no RDMA support, so GPU-to-GPU traffic would route through the host CPU. Clustering uses the two QSFP ports on the ConnectX-7 NIC, which run at up to 200 Gb/s with RoCE.
What cable connects two DGX Sparks, and what does it cost?
A single 0.5 m QSFP DAC, port-to-port, with no switch. At Resilient Tec the 200G QSFP56 MCP1650-V00AE30 is $79 (1–3 m versions $97–$145), and our NJAAKK-N911 / LMTQF022-SD-R compatible stacking cable, built to the NVIDIA-approved spec, is $174 (prices checked November 6, 2026).
Where can I buy the NJAAKK-N911 or MCP1650-V00AE30?
Directly from Resilient Tec, which ships from Knoxville, Tennessee, with POs and net terms available: NJAAKK-N911 — $174 at resilient-tec.com · MCP1650-V00AE30 — $79 at resilient-tec.com
What are the ports on the back of a DGX Spark for?
USB-C (4) for peripherals, storage and docks; one RJ-45 10GbE port for standard network and management traffic; two QSFP (ConnectX-7) ports at up to 200 Gb/s for GPU-to-GPU clustering, RDMA and high-throughput storage; and HDMI 2.1a for a local display.
Ready to Buy?
Tested and shipping from Knoxville, Tennessee. Lifetime warranty, backed by Resilient Tec. Prices below checked November 2, 2026.
- MCP1650-V00AE30 — 200G QSFP56 DAC, 0.5 m — $79, our best-selling Spark-to-Spark link
- MCP1650-V — 200G QSFP56 DAC, longer lengths — 1 m $97 · 1.5 m $104 · 2 m $115 · 3 m $145
- NJAAKK-N911 — QSFP112 400G stacking cable, 0.5 m — $174, compatible build to the Amphenol NJAAKK-0006 spec
- LMTQF022-SD-R — QSFP112 400G stacking cable, 0.5 m — $174, compatible build to the Luxshare LMTQF022-SD-R spec
Scaling to 3 or 4+ nodes, or mixing GB10 vendors? Send us your node count and hardware and we'll confirm exactly what to buy — including when a switch is the real answer.
You Might Also Need
- Can You Build a 4-Node DGX Spark Cluster Without a Switch? — switch vs. switchless ring at four nodes, and how to cable each
- Network switches — the supported route at four nodes and the only route past four; a 200GbE-capable managed switch, not a repurposed office switch
- QSFP-DD to QSFP56 breakout cables — the part that actually lets a switch talk to Spark-class ports at 4+ nodes
- Outgrowing desktop hardware? — how the same DAC-first approach scales up to rack-mounted DGX H100/H200 over NDR InfiniBand
- NVIDIA OSFP part-number crosswalk — for when the next step up is optics, not copper
Going Beyond the Spark
When a cluster outgrows the desk, the same ConnectX-7 fabric carries straight into rack hardware. Here is what we stock for the next step:
- MCP7H60 400G QSFP-DD to 2x200G breakout DAC — one 400G switch port feeds two Sparks, in lengths from 1 m to 3 m
- MikroTik CRS804 cluster cabling — a compact switch build for four or more GB10 nodes, and exactly which cables it takes
- Connect a Spark to a 100G switch, NAS or Jetson Thor — bringing the Spark into an existing network
- Which 400G ConnectX-7 optic? — MMA4Z00-NS400 vs MMA1Z00-NS400 vs MMS4X00-NS400, for runs too long for copper
- R740xd A40 GPU server — a rack-mounted, always-on inference server for team serving rather than a dev box; see what fits in 48 GB
- GPU servers for on-prem AI — our guide to refurbished Dell and HPE inference builds