NVIDIA DGX Spark Interconnect Compatibility Matrix
This guide maps the real documented interconnect options for the NVIDIA DGX Spark across ConnectX-7, Quantum-2, and Spectrum-X platforms. Specifications are sourced directly from NVIDIA's official DGX Spark documentation and networking platform pages (see sources below). Where NVIDIA has not published a specific validated pairing, we say so directly rather than guess.
→ See our related NVIDIA & Mellanox Interconnect Compatibility Guide
Part Number Library
| Component | Platform / Family | Part Number / Model | Key Specification |
|---|---|---|---|
| NVIDIA DGX Spark | Desktop AI Supercomputer | 940-54242-0000 | GB10 Grace Blackwell Superchip, 128GB unified memory, 2x QSFP56-DD ports (ConnectX-7) + 1x RJ-45 10GbE |
| ConnectX-7 | Integrated Smart NIC | Onboard (DGX Spark) | Dual QSFP56-DD ports on DGX Spark, 200Gb/s each, Ethernet mode; ConnectX-7 family (as a standalone HCA) supports up to 400Gb/s InfiniBand or Ethernet |
| Quantum-2 | InfiniBand NDR Switch Platform | QM9700 series | 64x 400Gb/s or 128x 200Gb/s ports, OSFP connectors, 51.2 Tb/s aggregate throughput |
| Spectrum-X | AI Ethernet Switch Platform | SN5000 / SN6000 series | Hyperscale AI Ethernet fabric switches, paired with Spectrum-X SuperNICs, RoCE-based |
Part Number Library — Extended NVIDIA Networking Portfolio
This section extends the library to NVIDIA's broader 400G–800G-class DPU, SuperNIC, and switch portfolio, sourced from NVIDIA's official networking documentation portal. These are separate product families from the DGX Spark's onboard ConnectX-7, which tops out at 200Gb/s per port — see the note below the table.
| Component | Platform / Family | Part Number / Model (OPN) | Speed & Connector | NVIDIA Lifecycle Status |
|---|---|---|---|---|
| NVIDIA BlueField-3 DPU | Data Processing Unit | B3240 — 900-9D3B6-00CN-AB0 | 2x QSFP112, InfiniBand NDR 400Gb/s / 400GbE, 16 Arm cores, PCIe Gen5 x16 | Mass Production |
| NVIDIA ConnectX-8 SuperNIC | InfiniBand/Ethernet SuperNIC | C8180 — 900-9X81E-00EX-ST0 | Single-cage OSFP, 800Gb/s XDR InfiniBand or 2x400GbE, PCIe Gen6 x16 | Mass Production |
| NVIDIA ConnectX-9 SuperNIC | InfiniBand/Ethernet SuperNIC | C9180 — 900-9X91E-00EB-ST0 | Single-cage OSFP, 800Gb/s XDR InfiniBand / 800GbE (NVIDIA markets up to 1.6Tb/s in dual-port configurations) | Prototype (not yet Mass Production) |
| NVIDIA Quantum-X800 | InfiniBand XDR Switch Platform | Q3400-RA — 920-9B36F-00RX-8S0 | 144 XDR ports over 72 OSFP cages, Quantum-3 ASIC, 800Gb/s per port | P-Rel (Pre-Release) |
| NVIDIA Spectrum-X (Spectrum-4 gen.) | AI Ethernet Switch Platform | SN5610 — 920-9N42F-00RI-3C1 | 64x OSFP ports, 800GbE per port, Spectrum-4 ASIC | P-Rel (Pre-Release) |
Note: NVIDIA BlueField-4 was announced in 2026 marketing materials describing 800Gb/s+ connectivity, but as of this writing NVIDIA's documentation portal has not published a BlueField-4 ordering-part-number guide, so we're not listing a specific model or OPN for it here rather than inventing one. We stock two real 800G OSFP breakout DAC cables that match the connector family used by ConnectX-8/9 and Quantum-X800/Spectrum-4 — MCP7Y00-N002 (800G OSFP → 2x400G OSFP) and MCP7Y50-N001 (800G OSFP → 4x200G OSFP). Neither is required for, or compatible with, the DGX Spark's own QSFP56-DD ports.
Compatibility Matrix
| Connection Path | Port / Connector | Speed | Officially Validated by NVIDIA? | Notes |
|---|---|---|---|---|
| DGX Spark ↔ DGX Spark (direct, 2-node) | QSFP56-DD ↔ QSFP56-DD | 200Gb/s | Yes | NVIDIA's official "Spark Stacking" guide documents this direct link for AI models up to 405B parameters. Ethernet mode only, not InfiniBand. We stock the matching 0.5m 200G QSFP56 DAC cable. |
| DGX Spark ↔ Generic QSFP Switch (4+ nodes) | QSFP56-DD | 200Gb/s per port | Generic requirement only | NVIDIA's "Multi Sparks Through a Switch" playbook specifies only "a QSFP switch with at least 4 QSFP56-DD ports (200Gb/s each)." No specific Quantum or Spectrum model is named. |
| DGX Spark ↔ NVIDIA Quantum-2 Switch | QSFP56-DD (Spark) to OSFP (switch) | 200Gb/s | Not named by NVIDIA | Quantum-2 switches use OSFP connectors, not QSFP. An OSFP-to-QSFP56 breakout cable (e.g. MFA7U10-H003) would be required; DGX Spark is not listed as a validated Quantum-2 endpoint in current NVIDIA documentation. |
| DGX Spark ↔ NVIDIA Spectrum-X Switch | QSFP56-DD (Spark) to Spectrum-X switch port | 200Gb/s | Not named by NVIDIA | Spectrum-X is marketed for hyperscale multi-GPU-server fabrics. NVIDIA's DGX Spark documentation does not list Spectrum-X as a validated pairing. |
Frequently Asked Questions
Does NVIDIA name a specific switch model for DGX Spark clusters?
No. NVIDIA's official multi-node playbook only specifies a generic requirement — a managed switch with at least 4 QSFP56-DD ports at 200Gb/s each. No specific Quantum-2 or Spectrum-X model is named as validated for DGX Spark clustering.
Can a DGX Spark connect directly to a Quantum-2 InfiniBand switch?
Physically, you'd need an OSFP-to-QSFP56 breakout or adapter cable, since DGX Spark uses QSFP56-DD ports while Quantum-2 switches use OSFP connectors. This path is not called out as an officially validated configuration in NVIDIA's current DGX Spark documentation, and the Spark's ConnectX-7 ports run in Ethernet mode, not InfiniBand.
What cable connects two DGX Spark units directly?
NVIDIA's documentation calls for an approved QSFP/ConnectX-7 cable between the matching ports on each unit (for example, both left ports or both right ports). NVIDIA currently supports stacking up to 2 units this way.
Why do some listings reference a "400G QSFP112" cable for DGX Spark when NVIDIA's spec sheet lists 200Gb/s QSFP56-DD ports?
NVIDIA's official DGX Spark hardware documentation describes two rear ConnectX-7 ports as QSFP56-DD running at 200Gb/s each, with no separate 400G port listed. If you come across a "400G QSFP112" cable marketed for DGX Spark, confirm the exact port and form factor against your unit's spec sheet before ordering — this detail isn't corroborated by NVIDIA's published specifications.
Do ConnectX-8, ConnectX-9, Quantum-X800, or Spectrum-4 connect directly to a DGX Spark?
No. The DGX Spark's onboard ConnectX-7 ports run at 200Gb/s over QSFP56-DD. ConnectX-8/9 SuperNICs and the Quantum-X800/Spectrum-4 switch families use OSFP connectors rated up to 800Gb/s — a separate, higher tier of NVIDIA's networking stack aimed at DGX B200/GB200-class and hyperscale AI factory deployments, not DGX Spark.
Is NVIDIA ConnectX-9 available to order today?
Per NVIDIA's own ordering-part-number tables (checked July 2026), every listed ConnectX-9 SKU carries a "Prototype" lifecycle status rather than "Mass Production." We're flagging this rather than presenting ConnectX-9 as generally available.
Is the Quantum-X800 switch in full production?
NVIDIA's Q32xx/Q34xx XDR switch ordering guide lists the Q3200-RA, Q3400-RA, and Q3401-RD models as "P-Rel" (Pre-Release) and the liquid-cooled Q3450-LD as "Prototype." None are yet marked "Mass Production" in NVIDIA's published documentation as of this writing.
Sources: NVIDIA DGX Spark product page and datasheet, DGX Spark User Guide (docs.nvidia.com), DGX Spark "Spark Stacking" and "Multi Sparks Through a Switch" playbooks (build.nvidia.com), NVIDIA Quantum-2 / Spectrum-X platform pages (nvidia.com/networking), and NVIDIA's official networking documentation portal (networking-docs.nvidia.com) for BlueField-3, ConnectX-8/9, and Quantum-X800/Spectrum-4 ordering part numbers. Last verified July 2026.