DGX Spark and GB10 cable finder

DGX Spark Cable Finder: The Right Cable and Count for Your Cluster

Every GB10 workstation uses the same 0.5m QSFP112 400G passive DAC, whichever of the eight vendors built it. Two units need one, and three or four units wired without a switch need one per unit. This DGX Spark cable finder gives the exact count and the wiring: choose your workstation and unit count below. Petronella Technology Group, Inc. stocks this part in the United States as the Amphenol NJAAKR-0006.

You need 1 cable.

Two NVIDIA DGX Spark units link directly with one 0.5m QSFP112 400G passive DAC, Port 0 to Port 0. No switch and no transceivers.

See the single cable: $159, in stock

Free US shipping, ships in 1 to 3 business days, US addresses only. International orders are quoted separately through the cable quote form.

DGX Spark cable finder results: cables by cluster size

Every GB10 system has exactly two QSFP112 ports on its ConnectX-7 network card, so the wiring is the same no matter which vendor built the box. That makes the count easy to state in one table.

UnitsCablesTopologySwitch needed?What to order
21Direct link, Port 0 to Port 0NoSingle cable, $159
33Ring: A to B, B to C, C to ANo3-pack ring kit, $435
44Ring: each unit's Port 0 to the next unit's Port 1No, optional4-pack four-node ring kit, $568
5 to 7One per unit, plus breakoutsSwitched fabricYesCall (919) 348-4912 for a port map
8One per unit, plus breakoutsSwitched fabricYesCall (919) 348-4912 for a port map

A spare is cheap insurance for a pair, which is why the 2-pack exists at $299. All four options, with the live stock status, are on the DGX Spark cluster cable page.

Order the cable or a ring kit

Which cable fits which GB10 workstation

The NVIDIA GB10 Grace Blackwell platform ships under eight brand names, and each vendor describes the interconnect a little differently. Some sell a first-party cable under their own part number, and some name none at all. Underneath, all eight carry the same ConnectX-7 network card with the same two QSFP112 ports, so one cable covers every one of them and brands can be mixed in a single cluster.

WorkstationThe vendor's own cableWhat fitsFit notes
NVIDIA DGX SparkNJAAKK-N911, the 400 mm stacking cable in NVIDIA's user guide0.5m QSFP112 400G passive DACDGX Spark cable
Dell Pro Max with GB10Dell 470-BDXP, listed by resellers at 25 cm and 200 Gb/s0.5m QSFP112 400G passive DACDell Pro Max GB10 cable
ASUS Ascent GX10ASUS sells its own QSFP cable0.5m QSFP112 400G passive DACASUS Ascent GX10 cable
MSI EdgeXpert MS-C931MSI accessory 956-C931-101, 400 mm0.5m QSFP112 400G passive DACMSI EdgeXpert cable
Lenovo ThinkStation PGXLenovo 4X91U42988, 0.4 m, up to 200 Gbps0.5m QSFP112 400G passive DACThinkStation PGX cable
HP ZGX Nano G1nHP option kit C20MQAA, a 0.4m QSFP112 DAC0.5m QSFP112 400G passive DACHP ZGX Nano cable
Acer Veriton GN100None named on our Acer page0.5m QSFP112 400G passive DACAcer Veriton GN100 cable
Gigabyte AI TOP ATOMGigabyte ATC9000-QSFP, 0.4 m, 30 AWG0.5m QSFP112 400G passive DACGigabyte AI TOP ATOM cable

If you are connecting ConnectX-7 adapter cards rather than GB10 workstations, the same part applies; see the ConnectX-7 QSFP112 DAC cable page.

Why one cable covers every brand

NVIDIA's Spark Stacking documentation names the approved build for these ports: the Amphenol NJAAKK0006, with the Luxshare LMTQF022-SD-R as the equivalent. The cable we stock is built to that same spec in the 0.5m length. Each one arrives factory sealed, with Amphenol AHSP P/N NJAAKR-0006, Rev G stamped on the jacket. It is passive copper with no firmware and no vendor lock, which is why the logo on the front of the workstation does not matter.

A passive direct-attach copper cable has no retimers, no digital signal processing, and nothing to update. It either links on day one or it does not, and once it links it stays linked. The cable is rated to 400G while the node-to-node link negotiates at 200 Gb/s, so it is over-provisioned by design. The ConnectX-7 card and the GB10's PCIe lanes set the ceiling, and the cable carries what they negotiate with margin to spare.

Factory-sealed Amphenol NJAAKR-0006 0.5m QSFP112 DGX Spark cable from our shelf stock

Our own shelf stock, photographed in-house: genuine Amphenol, jacket-stamped AHSP P/N NJAAKR-0006.

DGX Spark cable part numbers, decoded

Most of the confusion around this cable comes from part numbers that look alike. These facts come from NVIDIA's documentation and retailer listings, plus owner reports on the NVIDIA developer forum, checked September 2026.

Part numberLengthWhat it is
Amphenol NJAAKK-N911400 mmNVIDIA-approved. NVIDIA's DGX Spark user guide lists it as "QSFP to QSFP112, 32AWG, 400mm, LSZH." Retail listings for NVIDIA's own stacking cable carry this number.
Amphenol NJAAKK00060.5 mNVIDIA-approved. The same guide calls it "the 0.5m version" of the NJAAKK-N911.
Luxshare LMTQF022-SD-R400 mmNVIDIA-approved. Listed as "QSFP112 400G DAC Cable, 400mm, 30AWG."
Amphenol NJAAKR-00060.5 mA 0.5 m QSFP112 cable in the same Amphenol family, 30 AWG. This is the part printed on our current stock.
Lenovo 4X91U429880.4 mLenovo's first-party ThinkStation PGX QSFP Link Cable, listed with data rates up to 200 Gbps.
Dell 470-BDXP25 cmDell's first-party cluster cable for the Pro Max with GB10, listed by resellers at 200 Gb/s.

Our cable is built to the same NVIDIA spec as the NJAAKK-N911 and the NJAAKK0006, but it is not the same part number, and we do not sell it as one. The full comparison, including what owners report on the forum, is in our post on NJAAKK-N911 vs NJAAKK0006 part numbers.

How to wire each cluster size

Two units: one cable

Seat one end of the cable in a QSFP112 port on the first unit and the other end in a QSFP112 port on the second, pressing until the latch clicks. The connector is keyed and cannot be inserted upside down. Port 0 to Port 0 is the convention our vendor pages use, and either of the two ports on each machine works for a two-node link. There is no switch, no transceiver, and no optics to install, because the connector ends of a direct-attach cable are the transceivers. NVIDIA documents the cabling and the network setup in its Connect Multiple DGX Sparks for Distributed Workloads playbook, which covers two to four devices.

Three units: a switchless ring with three cables

Label the machines A, B, and C. Run one cable from A to B, a second from B to C, and a third from C back to A. Every box now uses both of its QSFP112 ports, and every pair of nodes has a direct link. That gives the ring full pairwise connectivity. Our port-by-port wiring guide for a three-Spark ring walks through the addresses and the checks.

Four units: a ring with four cables, or a switch

Four units can also run as a switchless ring: four cables, with each unit's Port 0 going to the next unit's Port 1. A ring of four is not a full mesh, but tensor-parallel inference only passes data between neighbors. In its October 22, 2026 eight-node results, NVIDIA calls the ring the simplest, lowest-cost four-node inference setup and measured it at 32.82 tokens per second of single-stream decode on Qwen3.8-27B, against 36.34 on a switch.

A switch costs more, but it lets you split the cluster into pairs or run any-to-any workloads. It also leaves room to grow. We compare both options in four DGX Sparks: switch or switchless ring.

Get the 3-pack or 4-pack ring kit

Five to eight units: plan a switched fabric

For eight nodes, a split cluster, or all-to-all workloads, use a 200G-class QSFP switch. Eight is the size NVIDIA now officially supports. Going through a switch changes the shopping list to one cable per node plus breakouts and the switch itself.

It also raises questions worth answering before hardware arrives:

  • How many ports do you need today, and how many after the next budget cycle?
  • Do the switch's port speed and form factor match the cables you already own?
  • Where does the switch physically sit relative to a stack of desk-side machines?

Petronella Technology Group, Inc. published its full write-up of a six-node GB10 fabric on two MikroTik CRS812 switches, with the switch configuration, the measured RDMA matrix, and vLLM tensor-parallel results, in the GB10 cluster guide on GitHub. The walk-through is also on our blog: from two Sparks to a switched fabric.

Is the cable in the box, and where are the ports?

Usually not, which is why this is the part that surprises buyers. HP's datasheet states that connecting two ZGX Nano systems requires a compatible QSFP cable, sold separately; its QuickSpecs list option kit C20MQAA, so a cable may appear on your order if it was configured at purchase. Lenovo's PSREF sheet for the ThinkStation PGX says appropriate ConnectX-7 cable options are required for every connection configuration, and Gigabyte sells its AI TOP QSFP cable separately from the AI TOP ATOM. If your order confirmation has no cable line item, you need one.

The two QSFP112 cages sit on the rear panel of every GB10 system. On the ASUS Ascent GX10 they are beside the 10G LAN jack. On the MSI EdgeXpert, Port 0 is the cage beside the RJ-45 jack and Port 1 is the cage beside the Kensington lock slot. Port numbering matters for the three-node and four-node rings, where each unit uses both ports, and it does not matter for a simple pair.

QSFP112 or QSFP56: which one does a DGX Spark need?

This is the most common question on the NVIDIA developer forums, and the confusion is understandable. The ConnectX-7 card in every GB10 system negotiates its node-to-node link at 200 Gb/s, which is inside QSFP56 territory, yet NVIDIA's Spark Stacking documentation specifies a QSFP112 400G DAC. Both statements are true at once. The QSFP112 part is the documented, approved cable, rated with headroom above what the link negotiates. Cable vendors and community testers have also reported 200G QSFP56 cables linking on DGX Spark pairs, because the port speaks the slower rate natively.

Our advice is simple: for a machine at this price, buy the cable NVIDIA documents. A QSFP112 400G DAC is guaranteed against the spec the Spark Stacking guide names, works identically on every GB10 build, and removes the one variable you would otherwise be debugging if the cluster link ever misbehaves. The connector is the standard QSFP form factor either way, which is why retail listings describe the official part as "QSFP/CX7": a QSFP-family cable for the ConnectX-7 card.

0.4 m or 0.5 m: picking a DGX Spark cable length

All of these cables fit the same port, so the choice comes down to length and timing. The 0.4 m parts suit units stacked directly on each other. The 0.5 m parts give a few more inches for side-by-side units and for the longest run in a three-node ring. Half a meter reaches between two stacked units or two neighbors on a bench with a clean service loop and no excess copper to coil.

Shorter is better with passive copper: less length means less signal loss and an easier time holding the negotiated rate. If your layout separates the machines, for example across a rack or through a desk grommet, tell us the actual port-to-port distance and we will quote the correct longer build rather than guessing.

What a cluster actually buys you

It helps to be clear about what the cable is for. Linking nodes pools their memory so you can load a model too large for a single 128 GB machine, and it lets a workload span multiple GPUs. It is a capacity play, not a raw speed boost, because each link runs at roughly 25 GB/s. For memory-bound inference and fine-tuning of larger models, that is exactly the win you want. For a job that already fits in one node, a second box adds headroom rather than throughput.

We measured 196 Gb/s across a single 0.5m cable in our own GB10 cluster work, and the results are public in the GB10 cluster guide, including a benchmark data chapter. For the deeper numbers, read what 400G really means for DGX Spark cluster bandwidth, and for what the pooled memory runs, see our LLM benchmark results.

Bringing the link up: a short checklist

NVIDIA's DGX Spark clustering documentation walks through the bring-up, and it is short:

  1. Seat both ends of each cable until the latch clicks.
  2. Confirm the ConnectX-7 interfaces are visible on each node.
  3. Assign each high-speed interface a static IP address on the same private subnet.
  4. Verify the link with a ping across those addresses.
  5. Point your distributed inference or training framework at the direct connection.

Budget fifteen minutes for the whole exercise, most of it reading rather than doing. If the link does not come up, the checklist is short precisely because the cable is passive. Reseat both ends and recheck the interface addresses. Then confirm both nodes are on current firmware. There is no cable firmware to suspect and no transceiver compatibility matrix to debug.

One practical buying note: the stacking cable is the piece that most often holds up an otherwise complete cluster. If your second workstation is on order, put the cable on order the same day.

Buying from Petronella Technology Group, Inc.

The cable is in stock in the United States and ships in 1 to 3 business days with free US shipping.

  • Single cable: $159
  • 2-pack: $299
  • 3-pack ring kit: $435
  • 4-pack four-node ring kit: $568

Checkout ships to US addresses only, and international orders are quoted separately. If your organization buys on purchase orders, or you need five or more cables, use the quote form on the cable page or call (919) 348-4912 with your quantity and ship-to address.

Petronella Technology Group, Inc. runs its own GB10 cluster, so the wiring advice on this page comes from machines on our own bench. If you are planning a larger build, we can map the ports and the switch model with you before you commit to any of it. Browse the rest of our AI hardware, or contact our team with your node count.

Order the DGX Spark cable Ask about a larger cluster

Frequently asked questions

What cable do I need to connect two DGX Sparks?

One 0.5m QSFP112 400G passive direct-attach copper cable, plugged into a QSFP112 port on each unit. NVIDIA's Spark Stacking documentation names the Amphenol NJAAKK0006 and Luxshare LMTQF022-SD-R spec, and our cable is built to it. There is a longer answer in what cable do I need to connect two DGX Sparks.

How many cables do I need for three DGX Sparks?

Three. A three-node cluster is a switchless ring: A to B, B to C, and C back to A. Each unit uses both of its QSFP112 ports. The 3-pack ring kit holds the three cables the ring needs.

Can four DGX Sparks run without a switch?

Yes. Four units can run as a switchless ring with four cables, each unit's Port 0 to the next unit's Port 1. It is not a full mesh, but NVIDIA's October 22, 2026 results call it the simplest, lowest-cost four-node inference setup. Choose a switch instead if you want to split the cluster into pairs or run all-to-all workloads. A switch is also the path to eight.

Does the same cable work on Dell, ASUS, MSI, Lenovo, HP, Acer, and Gigabyte GB10 systems?

Yes. Every GB10 workstation carries the same ConnectX-7 network card with two QSFP112 ports, so the cable is identical across all eight brands, and brands can be mixed in one cluster.

Is your cable the NVIDIA NJAAKK-N911?

No. The NJAAKK-N911 is a 400 mm Amphenol cable. Ours is the 0.5m length, jacket-stamped Amphenol NJAAKR-0006, built to the same NVIDIA Spark Stacking spec. It is an alternative to the NJAAKK-N911, not the same part number.

Will a cheaper QSFP56 cable work?

Cable vendors and community testers have reported 200G QSFP56 cables linking on DGX Spark pairs, because the port negotiates at 200 Gb/s. NVIDIA documents a QSFP112 400G DAC, and that is the cable we recommend, because it removes one variable if the link ever misbehaves.

How fast is the link between two units?

The ConnectX-7 card negotiates the node-to-node link at 200 Gb/s. We measured 196 Gb/s across a single 0.5m cable in our own cluster work. The cable is rated to 400G, so it is not the bottleneck.

Do I need a switch for eight units?

Yes. For eight nodes, use a 200G-class QSFP switch. The parts list becomes one cable per node plus breakouts and the switch itself. Call (919) 348-4912 and we will map the ports with you.

Do you ship outside the United States?

Checkout ships to US addresses only. International orders are quoted separately through the quote form on the cable page.

Written by Craig Petronella, founder of Petronella Technology Group, Inc. Craig holds the CMMC-RP, CCNA, and CWNE credentials and is a licensed North Carolina Digital Forensic Examiner (604180-DFE). His team built and measured the GB10 cluster this page draws on.

Last reviewed November 19, 2026.

Sources