The NVIDIA DGX Spark is already a remarkable machine on its own — a desktop-sized AI supercomputer packing up to 1 petaFLOP of FP4 AI performance, 128 GB of unified memory, and a full NVIDIA software stack. But there are very real reasons why researchers, developers, and enterprises are looking to link two or more of them together. This post breaks down the why, the hardware limitations, and exactly how far you can scale a multi-Spark cluster.
Short answer: the cable and what it costs
Two DGX Sparks connect with one 0.5 m passive DAC — $79 or $174 at Resilient Tec
Plug a single 0.5 m QSFP-to-QSFP passive DAC between the ConnectX-7 ports of the two units — no switch needed. NVIDIA validates the Amphenol NJAAKK-N911 and Luxshare LMTQF022-SD-R (both QSFP112). Because the Spark’s ConnectX-7 ports negotiate at up to 200 Gb/s, the lower-cost QSFP56 MCP1650-V00AE30 also links reliably.
|
MCP1650-V00AE30 — QSFP56 200G DAC, 0.5 m Lowest-cost route. Works reliably on DGX Spark. |
$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. A three-node ring needs three cables; four nodes take a 200 GbE switch (the supported route) or an unsupported switchless ring of four cables. The MCP1650-V also comes in 1–3 m lengths ($97–$145) for switch runs and spread-out units. Prices checked November 9, 2026. POs and net terms available.
Why Would You Connect Two DGX Sparks?
The DGX Spark is purpose-built to handle large AI workloads at your desk — fine-tuning models up to 70 billion parameters, running inference on models up to 200 billion parameters, and powering autonomous AI agent pipelines. But some workloads quickly push against the limits of a single node, and that is precisely where linking two units together becomes compelling.
1. Running Larger Models
A single DGX Spark has 128 GB of coherent unified system memory. That is enough to run inference on models up to roughly 200 billion parameters and fine-tune up to 70 billion parameters. Once you connect a second unit, you effectively double the available memory pool, unlocking support for models in the 400B-parameter range that simply cannot fit on one node.
2. Faster Fine-Tuning and Training
Fine-tuning workloads can be parallelized across nodes with near-linear performance scaling. NVIDIA's own benchmarks show that going from one to two DGX Spark nodes for fine-tuning a Nanochat model doubles total token throughput from approximately 18,400 tokens/second to 35,900 tokens/second. For teams iterating frequently on custom models, this halving of fine-tuning time is a significant productivity gain.
3. Faster Inference via Tensor Parallelism
For latency-sensitive agentic workloads, spreading inference across two nodes using tensor parallelism can meaningfully reduce time-to-first-token (TTFT). Running Llama 3.3 70B on two nodes drops TTFT from around 33,415 ms to 21,384 ms — a roughly 36% improvement — and cuts per-token generation time approximately in half.
4. Running More Agents in Parallel
Autonomous AI agent frameworks — such as those built on NVIDIA’s NeMo Agent toolkit — often require multiple simultaneous subagents, each managing long context windows. A single DGX Spark can parallelize several subagents across its GPU, but distributing agent workloads across two nodes increases concurrency capacity and total throughput for multi-agent pipelines.
5. Prototyping at Scale Before Cloud Deployment
A two-Spark setup gives developers a local multi-node environment that mirrors cloud-scale distributed inference and training workflows. You can develop, test, and validate multi-node workloads on-premises — with full data privacy — before deploying to larger cloud infrastructure.
Hardware Limitations: What You Need to Know
The DGX Spark's interconnect capability is built around its NVIDIA ConnectX-7 NIC, which runs at 200 Gbps. This is the single most important hardware constraint to understand when planning a multi-Spark setup.
The ConnectX-7 NIC: Your Bandwidth Budget
Each DGX Spark ships with one ConnectX-7 NIC operating at 200 Gbps. This port supports RDMA over Converged Ethernet (RoCE), which is the protocol used to enable low-latency, GPU-direct communication between nodes. In a two-node direct cable connection, you are working with a single 200 Gbps link between the two machines.
By comparison, a server-class DGX H100 system uses NVLink Switch connections for terabytes per second of inter-GPU bandwidth. The DGX Spark's 200 Gbps Ethernet interconnect is capable for many workloads, but it is a real ceiling for communication-intensive tasks like large-model tensor parallelism.
Memory Bandwidth Per Node
Each DGX Spark delivers 273 GB/s of memory bandwidth from its 128 GB LPDDR5x unified memory. This is substantially lower than the HBM3e memory found in data center Blackwell GPUs, which is why intra-node memory-bound workloads are more constrained than on cloud-scale hardware. When fine-tuning or running inference across two Spark nodes, the interconnect bandwidth between the two machines becomes a secondary bottleneck if the model requires frequent synchronization.
Power Per Unit
Each DGX Spark draws up to 240 watts (with the GB10 chip at a 140 W TDP). A two-unit cluster draws around 480 watts total — a factor to consider for home office or lab deployment, though well within reach of standard power infrastructure.
Scaling Is Not Always Linear
Inference scaling across nodes is near-linear only when inter-node communication is minimal — for example, when running many independent simulations or data-parallel workloads. For communication-intensive tasks like large tensor-parallel inference, scaling is sub-linear because the 200 Gbps NIC becomes a bottleneck relative to the volume of data that must be synchronized between GPUs.
Can You Link More Than Two? Yes — Up to Eight
NVIDIA now officially supports scaling DGX Spark up to eight nodes (announced October 22, 2026). NVIDIA Sync still sets up as many as four through a switch; past that you configure by hand. The platform supports multiple execution topologies, each suited to different workload profiles:
Two-Node Direct Cable Connection
The simplest setup: connect two DGX Sparks back-to-back using a direct QSFP112 400G-rated DAC cable between their ConnectX-7 QSFP ports (also marketed as a QSFP/CX7 cable; each port runs at 200 Gb/s). The two cables NVIDIA approves for this are the NJAAKK-N911 and the Luxshare LMTQF022-SD-R (Resilient Tec stocks an LMTQF022-SD-R compatible build), both 0.5m passive direct-attach copper. Resilient Tec sells the NJAAKK-N911 and LMTQF022-SD-R compatible for $174 each, and the lower-cost 200G QSFP56 MCP1650-V00AE30 for $79 — because the Spark’s ConnectX-7 ports negotiate at up to 200 Gb/s, it also links reliably. If the two units are not side by side, the same cable comes in 1 m and longer. No switch is required. This topology is ideal for balanced scaling on larger models (up to ~400B parameters), faster fine-tuning, and general-purpose distributed inference. Setup is straightforward and latency is excellent due to the point-to-point connection.
Three-Node Ring Topology
Three DGX Sparks can be connected in a ring, where each node connects to the next in a loop. This topology is particularly well-suited for fine-tuning larger models or running smaller training jobs. Ring configurations distribute collective communication — like all-reduce operations used in distributed training — efficiently across the three nodes. NVIDIA provides a dedicated playbook for this setup. A ring uses three cables, one per link — three compatible NJAAKK-N911 cables come to $475 at Resilient Tec as a ring kit ($174 each separately), or three MCP1650-V00AE30 cables to $237 ($79 each).
Four-Node Cluster with a 200 GbE Switch
NVIDIA's supported way to connect four DGX Sparks is a 200 GbE network switch. All four nodes connect through the switch, which enables full any-to-any communication between all nodes. Cable it with one 200G QSFP56 DAC per Spark — the MCP1650-V in 1 m ($97), 1.5 m ($104), 2 m ($115) or 3 m ($145) — or with QSFP-DD to 2×200G breakout cables when the switch has fewer 400G ports than nodes. This is the highest-performance multi-Spark configuration currently supported, enabling local inference on state-of-the-art models up to 700 billion parameters and communication-intensive workloads. Four-node benchmarks show near-linear fine-tuning scaling — for example, Isaac Lab reinforcement learning FPS scales from 630 on one node to 2,520 on four nodes, and Nanochat fine-tuning throughput reaches ~74,600 tokens/second across four nodes versus ~18,400 on a single node.
Four-Node Switchless Ring (Unsupported)
Owners have also run four Sparks with no switch, as a ring: four cables, each unit using both of its QSFP ports to reach its two neighbors. It works for tensor-parallel inference because NCCL's ring all-reduce only passes data between neighboring nodes — but it needs a custom NCCL build and topology plan, NVIDIA does not support it, and it cannot grow past four nodes. Four NJAAKK-N911 cables come to $696 at Resilient Tec, or four MCP1650-V00AE30 cables to $316. Measured numbers, rack cabling and ring-vs-switch trade-offs →
Which Topology Is Right for You?
Here is a quick summary of how the configurations compare:
| Configuration | Hardware Needed (Resilient Tec price) | Best For | Max Model Size |
|---|---|---|---|
| 1 DGX Spark | — | Large context inference, fine-tuning up to 70B, local agent development | ~200B parameters |
| 2 DGX Sparks (direct) | 1 × 0.5 m DAC: NJAAKK-N911 ($174) or MCP1650-V00AE30 ($79) | Faster fine-tuning, larger models, low-latency inference | ~400B parameters |
| 3 DGX Sparks (ring) | 3 × NJAAKK-N911 ($174 each) or 3 × MCP1650-V00AE30 ($79 each) | Fine-tuning larger models, small training jobs | ~500B+ parameters |
| 4 DGX Sparks (switch) | 200 GbE switch + 4 × MCP1650-V DAC (1–3 m, $97–$145 each) or 2 × QSFP-DD breakout | State-of-the-art model inference, communication-intensive workloads | ~700B parameters |
| 4 DGX Sparks (switchless ring, unsupported) | 4 × NJAAKK-N911 ($174 each) or 4 × MCP1650-V00AE30 ($79 each) + custom NCCL | Tensor-parallel inference when a switch isn't available | ~700B parameters |
The Bottom Line
Connecting two DGX Sparks together makes compelling sense the moment your models or workloads outgrow a single node's 128 GB of memory or you need to cut fine-tuning times roughly in half. The hardware foundation — the ConnectX-7 NIC at 200 Gbps per node — is purpose-built for this, and the two-unit direct-cable setup — using an approved QSFP112 400G-rated DAC cable — requires no additional switching hardware at all. That one link costs $174 with our NJAAKK-N911-compatible cable or $79 with the MCP1650-V00AE30, both sold by Resilient Tec.
Beyond two units, NVIDIA's support extends to ring topologies (three nodes) and switch-based clusters of up to eight nodes (1 TB of unified memory), which opens frontier-class models that will not start on four. Four-node switch clusters remain what NVIDIA Sync sets up today. The main hardware constraints to keep in mind are the 200 Gbps per-node NIC bandwidth, the 128 GB per-node memory ceiling, and the sub-linear scaling on communication-heavy workloads. Owners are also running four nodes as a switchless ring — workable, but outside NVIDIA support.
For most AI development teams, a two-Spark setup hits a compelling sweet spot: no switch required, near-linear fine-tuning scaling, and access to 400B-parameter models — all on your desk, with full data privacy.
Interested in setting up a multi-Spark cluster? NVIDIA offers official playbooks for connecting multiple DGX Sparks, including step-by-step guides for ring topologies and switch-based four-node clusters, available at build.nvidia.com/spark.
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. Adds measured ring-vs-switch and breakout numbers to the topology choices above. See part two of our ring vs. switch benchmarks.
FAQ
What cable do I need to connect two NVIDIA DGX Sparks?
One 0.5 m QSFP-to-QSFP passive DAC between the two ConnectX-7 ports, with no switch. NVIDIA validates the Amphenol NJAAKK-N911 and Luxshare LMTQF022-SD-R (both QSFP112). Because the Spark’s ConnectX-7 ports negotiate at up to 200 Gb/s, the lower-cost 200G QSFP56 MCP1650-V00AE30 also links reliably.
How much does a DGX Spark stacking cable cost?
At Resilient Tec, our NJAAKK-N911-compatible cable is $174, an LMTQF022-SD-R compatible cable is $174, and the MCP1650-V00AE30 200G QSFP56 DAC is $79. All three are 0.5 m, sold new with a lifetime warranty (prices checked November 6, 2026). Longer MCP1650-V lengths run $97 (1 m) to $145 (3 m).
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
Do I need a switch to connect DGX Sparks, and how many cables?
Two nodes: no switch, one cable. Three nodes in a ring: no switch, three cables (three NJAAKK-N911 cables are $475 as a ring kit at Resilient Tec, or three MCP1650-V00AE30 cables $237). Four nodes: a 200 GbE switch is the supported route, with one MCP1650-V DAC per Spark ($97–$145 by length); a switchless ring of four cables ($696 in NJAAKK-N911, $316 in MCP1650-V00AE30) also works with custom NCCL software, outside NVIDIA support.
Do I need two cables between two DGX Sparks to get more bandwidth?
No. NVIDIA's Sync Cluster Assistant says to use only one QSFP cable per link, for direct and switch connections alike. One cable already carries the Spark's full 200 Gb/s per port, because each port is two 100G halves riding the same cable; owners report about 196 Gb/s on a single cable once both halves are driven. If you see about 100 Gb/s, the fix is configuration, not a second cable. The NJAAKK-N911 QSFP112 stacking cable is $174 at Resilient Tec, and the MCP1650-V00AE30 200G QSFP56 DAC is $79, both new with a lifetime warranty. (Prices checked November 9, 2026.)
Ready to Buy?
Tested and shipping from Knoxville, Tennessee. Lifetime warranty, backed by Resilient Tec. Prices below checked October 26, 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, for switch runs and spread-out units
- NJAAKK-N911 — QSFP112 400G stacking cable, 0.5 m — $174, compatible build to the Amphenol NJAAKK-0006 spec
- LMTQF022-SD-R compatible — QSFP112 400G stacking cable, 0.5 m — $174, Resilient Tec build cross-referenced to Luxshare LMTQF022-SD-R
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 past four (NVIDIA supports up to eight through a switch); 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
- Layer 1 Decoded: Every Physical Form Factor in a Modern Rack — QSFP56, QSFP112, QSFP-DD and OSFP side by side
- DAC vs AOC vs AEC — which cable goes where once your cluster outgrows the desk
- 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