64GB DGX Spark: How to Connect Two, and Can You Mix One With a 128GB Spark?

NVIDIA's 64GB DGX Spark goes on sale October 23, starting at $4,999, sold only through Acer, ASUS, Dell, Gigabyte, HP and MSI. It is the same GB10 machine with half the memory, and NVIDIA's pitch is that you buy two and cluster them to 128GB. Two 64GB Sparks connect with one QSFP cable between their ConnectX-7 ports. That's the same cable and setup as any other GB10 pair. Nothing in the launch coverage says the cable is in the box.

Below: which cable to buy, how the pair is set up, and the question the launch material skips: what happens when you pair a 64GB unit with a 128GB one.

What the 64GB DGX Spark Is (and Isn't)

Memory is the only difference the launch material describes. The cluster network is identical, so everything already known about clustering GB10 systems carries straight over: the cable, the ports, the RoCE setup and the topology limits.

Which Cable to Buy

NVIDIA's DGX Spark documentation names two approved cables, both QSFP112 passive copper: the Amphenol NJAAKK-N911 (400 mm, with NJAAKK0006 as its 0.5 m version) and the Luxshare LMTQF022-SD-R (400 mm). The same documentation caps each port at 200 Gb/s and says a faster cable brings no benefit, so a 200G QSFP56 DAC links the pair at the same speed for less:

Pick the QSFP112 cable if you want the cable type in NVIDIA's guide; pick the 200G DAC to save $70. The full comparison is in MCP1650-V00AE30 vs NJAAKK-N911.

Setting Up the Pair

1. Update both units to the same DGX OS and firmware before cabling. Partner firmware does not always land on the same day, and a firmware mismatch is the most common cause of a slow link.
2. Connect one cable between a QSFP port on each unit, pull-tab up. Check your partner's manual for which rear port is which; layouts differ by chassis.
3. Run NVIDIA Sync's Cluster Assistant. NVIDIA says it discovers the units, validates the configuration and sets up the ConnectX-7 network. NVIDIA's "Connect Two Sparks" playbook is the manual route.
4. Check the link. The NCCL bandwidth test should report close to 25 GB/s (200 Gb/s). About half that means only one of the port's two interfaces is carrying traffic; see GB10 stuck at 100 Gb/s.

The step-by-step guides for each brand are in How to Connect Two ASUS Ascent GX10s and How to Connect Two Dell Pro Max GB10 or HP ZGX Nano Systems. The 64GB versions cable exactly the same way.

Can You Mix a 64GB Spark With a 128GB Spark?

This is the question the launch coverage doesn't answer, and it splits into two parts.

Will they link? Yes. The 64GB and 128GB models have the same ConnectX-7 NIC and QSFP ports, so the same cable links them at 200 Gb/s. GB10 systems from different brands already cluster this way in the field. NVIDIA's 64GB material only describes pairing two 64GB units, so a mixed pair is outside what NVIDIA has described.

Will you get 192GB? Usually not. How much model a pair can hold depends on how the software splits it:

• Tensor parallelism is how vLLM and NVIDIA's cluster playbooks run a model across Sparks. It divides the model's weights evenly between the nodes. On a 64GB + 128GB pair, each node gets the same share, so the 64GB node sets the limit and the pair holds about what two 64GB units hold, roughly 128GB.
• Uneven splits are possible in some tools. llama.cpp, for example, can assign a larger share of a model's layers to a bigger node. That uses more of the 128GB unit, but it is a different software path from NVIDIA's playbooks, and speed depends on the slower share.

If you already own a 128GB Spark, a 64GB unit adds a second GPU and some memory for less money, but plan model sizes around two equal halves. If you are buying both units now, two of the same model is the simpler cluster.

Three 64GB Sparks

The 64GB model has two QSFP ports like every GB10 system, so three units can form the switchless three-cable ring NVIDIA documents for DGX Spark. That pools 192GB. NVIDIA's 64GB launch material only covers pairs, so treat a 64GB ring as the documented Spark ring rather than a 64GB-specific promise. The wiring is in How to Connect Three DGX Sparks. Three cables cost $462 in QSFP112 or $252 in 200G DAC.

Two 64GB Units or One 128GB Unit?

NVIDIA's own numbers make the case for the pair: on Qwen3.8 27B, two clustered 64GB units ran up to 1.7× faster than a single 128GB unit, because the pair has two GPUs working on the same model. The trade-offs:

• Same pooled memory. Both setups hold about the same model size.
• The pair adds a network hop. Each link runs at about 25 GB/s, roughly a tenth of the GB10's memory bandwidth. Workloads that shuffle a lot of data between nodes gain less than NVIDIA's example.
• The pair is two machines. Two sets of firmware to keep matched, and two power bricks, but also a second box you can split off for a different job.

FAQ

How do you connect two 64GB DGX Sparks?
With one QSFP passive copper cable between a ConnectX-7 QSFP port on each unit. The 64GB model has the same GB10 chip and the same two 200GbE QSFP ports as the 128GB model, so it uses the same cable and the same setup. NVIDIA Sync's Cluster Assistant discovers the pair and configures the ConnectX-7 network, and the two units pool 128GB of memory.

Does the 64GB DGX Spark come with a QSFP cable?
None of the launch coverage or the partner announcements say a cable is included, so plan on buying one. One cable links two units. Check the box contents on your specific partner model, because each of the six partners ships its own SKU.

Can a 64GB DGX Spark cluster with a 128GB DGX Spark?
On the network, yes: both have the same ConnectX-7 NIC and QSFP ports, so they link with the same cable at 200 Gb/s. NVIDIA's launch material only describes pairing two 64GB units, though. For model size, tensor parallelism, the method vLLM and NVIDIA's playbooks use, splits a model evenly across nodes, so a 64GB + 128GB pair holds roughly what two 64GB units hold. Tools that can split a model unevenly, such as llama.cpp, can put more of it on the bigger node.

Which cable connects two 64GB DGX Sparks, and what does it cost?
The NJAAKK-N911 compatible QSFP112 stacking cable, 0.5 m, is $154 at Resilient Tec; it matches the cable type NVIDIA names for DGX Spark. The MCP1650-V00AE30 200G QSFP56 DAC, 0.5 m, is $84 at Resilient Tec and links the pair at the same 200 Gb/s. Both are new with a lifetime warranty. (Prices checked October 19, 2026.)

Where can I buy a cable for the 64GB DGX Spark?
Directly from Resilient Tec, shipping from Knoxville, Tennessee, with POs and net terms available. The NJAAKK-N911 compatible stacking cable is $154 at resilient-tec.com and the MCP1650-V00AE30 200G DAC is $84 at resilient-tec.com. Both are in stock, new, with a lifetime warranty. (Prices checked October 19, 2026.)

Can I connect three 64GB DGX Sparks?
The hardware supports it: like every GB10 system, the 64GB model has two QSFP ports, so three units form a switchless ring with three cables, using the layout NVIDIA documents for three Sparks. That gives 192GB of pooled memory. NVIDIA's 64GB launch material only describes two-unit clusters, so treat three as the documented DGX Spark ring rather than a 64GB-specific claim.

Is two 64GB DGX Sparks better than one 128GB DGX Spark?
It depends on the model. NVIDIA reports two clustered 64GB units ran Qwen3.8 27B up to 1.7x faster than a single 128GB unit, because the pair has two GPUs. Both setups hold about the same model size. A single 128GB unit avoids the network hop, so models that need the whole 128GB on one GPU still favor it.

Ready to Buy?

Everything you need to cluster 64GB Sparks. New, lifetime warranty, shipping from Knoxville, Tennessee. Prices below checked October 19, 2026.

• NJAAKK-N911 compatible — QSFP112 400G stacking cable, 0.5 m — $154, in stock at Resilient Tec; one for a pair, three for a ring
• LMTQF022-SD-R compatible — QSFP112 400G stacking cable, 0.5 m — $154, the same cable under Luxshare's part number
• MCP1650-V00AE30 — 200G QSFP56 DAC, 0.5 m — $84, in stock at Resilient Tec; the same 200 Gb/s link for less
• MCP7H60-W002 — 400G QSFP-DD to 2×200G QSFP56 breakout, 2 m — $179.45, switch side only, one per pair of units at four or more

Ordering Sparks for a team? Send us your models and node count and we will confirm the cables before the units arrive. POs and net terms available.

You Might Also Need

• How to connect three DGX Sparks — the three-cable ring, port by port
• Can you cluster different GB10 systems? — mixing brands, and the four things that actually differ
• MCP1650-V00AE30 vs NJAAKK-N911 — which cable to choose
• How to connect two Dell Pro Max GB10 or HP ZGX Nano systems — including the firmware trap
• DGX Spark interconnect reference — part numbers and what to buy at each node count

Sources

• StorageReview — NVIDIA DGX Spark 64GB lands October 23 at $4,999
• Data Studios — NVIDIA launches 64GB DGX Spark with two-system clustering
• NVIDIA DGX Spark User Guide — ConnectX-7 networking and approved cables
• NVIDIA — Connect Three DGX Spark in a Ring Topology (playbook)
• vLLM documentation — parallelism and scaling