ConnectX-7 vs ConnectX-8 vs ConnectX-9: Which NVIDIA AI Network Card Fits Which Job?
Every GPU in an AI cluster talks to the others through a network card. For NVIDIA systems that card is almost always a ConnectX, and there are now three generations on sale or arriving: ConnectX-7, ConnectX-8 and ConnectX-9.
The numbers double each time, which makes it tempting to assume the newest is the one to buy. It usually is not a free choice. The card follows the GPU and the server, and the fastest card in the wrong server is money spent on speed it cannot use.
The three generations at a glance
| Compare | ConnectX-7 | ConnectX-8 SuperNIC | ConnectX-9 SuperNIC |
|---|---|---|---|
| 🚀 Top speed | 400 Gb/s |
800 Gb/s |
Up to 1.6 Tb/s per GPU |
| 🌐 Ethernet per port | 10 to 400 GbE | 25 to 400 GbE, typically as 2 × 400 GbE | 25 to 800 GbE |
| 🔗 InfiniBand | Up to NDR, 400 Gb/s | Up to XDR, 800 Gb/s | Up to 800 Gb/s per port |
| 🧱 Server slot it needs |
PCIe Gen5 Up to 32 lanes; also works in Gen4 servers |
PCIe Gen6 Up to 48 lanes; Gen5 compatible |
PCIe Gen6 Up to 48 lanes; Gen5 compatible |
| 🧠 GPU generation it was built for | Hopper-era systems, general servers and GB10 desktop AI systems | Blackwell and Blackwell Ultra | Rubin (Vera Rubin platform) |
| ⭐ What is new | The flexible all-rounder: one, two or four ports | A PCIe switch built into the card | 800 Gb/s on a single Ethernet port |
Sources: NVIDIA ConnectX-7 datasheet, PNY: NVIDIA ConnectX-8 SuperNIC, NVIDIA ConnectX-8 user manual, PNY: NVIDIA ConnectX-9 SuperNIC, NVIDIA ConnectX-9 firmware release notes. The "built for" and "what is new" rows are our own summary. Exact speeds depend on the model number; check the part you are quoted.
ConnectX-7: the all-rounder
ConnectX-7 supports both InfiniBand and Ethernet at up to 400 Gb/s, comes with one, two or four ports, and connects through PCIe Gen5. It works with Ethernet switches from 1 Gb/s to 400 Gb/s, and there are versions for servers that only have PCIe Gen4.
That breadth is the point. It will go into almost any current server and talk to almost any switch you already own.
Use it for:
- H100 and H200 clusters, where 400 Gb/s per GPU is the designed speed
- Inference servers built on PCIe cards such as the RTX PRO 6000 Blackwell
- Storage servers feeding a GPU cluster
- Upgrading an existing data centre network to 100, 200 or 400 GbE
- Traditional high-performance computing on InfiniBand
- Linking desktop AI systems together, which is the use most readers can try today
The ConnectX-7 use case on your desk: stacking two DGX Spark units
You do not need a data hall to use ConnectX-7. The NVIDIA DGX Spark and the other desktop systems built on the same GB10 chip, such as the ASUS Ascent GX10 and Lenovo ThinkStation PGX, have a ConnectX-7 with two 200 GbE QSFP ports on the back.
Join two units through those ports with a single cable and they behave as a two-node cluster. The memory of both is pooled, so a model too large for one 128 GB unit can run across the pair. NVIDIA calls this Spark Stacking.
| How many units | How they connect | What you need |
|---|---|---|
| 🖥️🖥️ Two |
Direct One cable between the same ConnectX-7 port on each unit, no switch |
One 200G QSFP56 DAC cable |
| 🖥️🖥️🖥️ Three or more, through a 200G switch |
Switch Each unit cabled to its own 200 GbE port |
A 200 GbE switch such as the NVIDIA Spectrum-2 SN3700, plus one QSFP56 DAC per unit |
| 🖥️🖥️🖥️🖥️ Three or more, through a 400G switch |
Switch + breakout Each 400G port split into two 200G links |
A 400G switch such as the MikroTik CRS804 DDQ, plus 400G-to-2×200G breakout cables |
Sources: NVIDIA developer forum: DGX Spark stacking, Layer3Labs: DGX Spark cluster setup and sizing, NADDOD: how to deploy NVIDIA DGX Spark. The products linked are examples from the SourceIT catalogue, not an NVIDIA-validated bill of materials.
Three practical notes before you order:
- Check the cable against NVIDIA's list. NVIDIA's Spark Stacking guide names specific approved cables. The cable we list is a compatible 200G QSFP56 DAC made for this job, in 1, 2 and 3 metre lengths. Confirm it suits your units before buying, and contact us if you are unsure.
- One cable is enough. NVIDIA's guidance, as quoted by users on its developer forum, is that a single QSFP cable gives the full 200 GbE between two units. Adding a second cable between the same pair does not double the speed.
- Stacking adds memory more than speed. Two units let you load a larger model. They do not make a small model answer twice as fast. Use our GPU sizing guide to check whether your model needs the second unit at all.
ConnectX-8: the card that is also a switch
ConnectX-8 doubles the speed to 800 Gb/s and moves to PCIe Gen6. NVIDIA calls it a SuperNIC, a term we explain in NIC, SuperNIC or DPU?
The more interesting change is inside. ConnectX-8 has up to 48 PCIe lanes and an integrated PCIe switch. In a Blackwell system it is not only the network card. It is also the part that connects GPUs, processors and storage drives to each other. NVIDIA presented it as the first deployment of a PCIe Gen6, 800 Gb/s SuperNIC, on the GB300 platform.
One detail matters for anyone planning an Ethernet network. The 800 Gb/s version runs as a single 800 Gb/s InfiniBand link, or as two 400 Gb/s Ethernet links. On Ethernet, each port tops out at 400 GbE.
Use it for:
- HGX B300 8-GPU servers, which are specified with eight ConnectX-8 SuperNICs. Examples: HPE Compute XD690, Dell PowerEdge XE9780 / XE9785, Lenovo ThinkSystem SR680a V4 and ASUS XA NB3I-E12
- GB300 NVL72 racks such as the NVIDIA GB300 NVL72 by HPE, where it is part of the design
- New InfiniBand clusters at 800 Gb/s
- Spectrum-X Ethernet fabrics for large-scale training
Sources: NVIDIA's ConnectX-8 presentation at Hot Chips 2025, ConnectX-8 firmware release notes, Introl on B300 infrastructure
ConnectX-9: built for Rubin
ConnectX-9 is the network card of the Vera Rubin platform. NVIDIA describes it as delivering up to 1.6 terabits per second of throughput to Rubin GPUs, with up to 800 Gb/s per port on both InfiniBand and Ethernet. It keeps PCIe Gen6 with up to 48 lanes, and it is designed to pair with NVIDIA's Spectrum-X Ethernet and Quantum-X800 InfiniBand switches.
The step that matters for Ethernet users is the port speed. Where ConnectX-8 offered 400 GbE per port, ConnectX-9 lists 800 GbE.
Use it for:
- Vera Rubin systems, where it is part of the platform
- New large clusters being designed around 800 GbE switching
For most other buyers today it is a card to know about, not one to order. See Vera Rubin NVL72 vs GB300: Should You Still Order?
Which card for which job
| Your project | Start with |
|---|---|
| 🖥️ Two to four desktop AI systems on GB10 | ConnectX-7 Already built in; you only need the cable or switch |
| 💬 One or two inference servers for internal users | ConnectX-7 At 100 or 200 GbE; more is rarely needed |
| 💾 Storage nodes for a GPU cluster | ConnectX-7 |
| 🗄️ An H200 training cluster | ConnectX-7 At 400 Gb/s |
| 🧠 HGX B300 servers or a GB300 rack | ConnectX-8 It comes with the platform |
| 🏭 A Vera Rubin system | ConnectX-9 It comes with the platform |
| 🔄 Upgrading cards in servers you already own | Check the slot first See below |
Three things the speed figure hides
1. The slot sets the ceiling
A network card can only move data as fast as the server can hand it over. A PCIe Gen5 slot with 16 lanes carries roughly 500 Gb/s by our arithmetic, which is why 400 GbE is described as a speed Gen5 slots can handle. An 800 Gb/s card needs PCIe Gen6, or a second connection to the motherboard through an auxiliary kit. Both newer cards are backwards compatible with Gen5, so they will work in an older server. They will not reach full speed there.
2. The switch has to match
A card is one end of a link. An 800 Gb/s card connected to a 100 Gb/s switch runs at 100 Gb/s. The newer cards do work with slower switches, which is useful during a staged upgrade, but the budget for a faster network is mostly switches, optics and cables, not cards.
3. The connector changes
The faster cards use different plugs. ConnectX-7 comes with QSFP112 or OSFP ports, and the 800 Gb/s cards use OSFP. Cables and transceivers are not always interchangeable between them. We cover the cable side in Copper, Optical or Co-Packaged?
Sources: ServeTheHome on ConnectX-7 and 400GbE, NVIDIA ConnectX-9 specifications, NVIDIA ConnectX-8 introduction
The slightly controversial view
For most enterprises, the network card is the last place to chase the newest generation. A team running a few inference servers will see no difference between 400 and 800 Gb/s, because the traffic is not there. Network speed matters when many GPU servers train one model together and spend their time exchanging results. Below that scale, a well-configured 100 or 200 GbE network on ConnectX-7 is the sensible purchase.
The counter-argument is worth stating. Networks outlive servers. Switches and cabling bought today will still be in the rack when the next two generations of GPU arrive, so there is a case for building the fabric one step ahead of the servers. That is a reasonable choice for a site that knows it will grow. It is an expensive one for a site that only hopes to.
Five questions for the quote
- Which ConnectX generation and exact model number is in this server, and how many?
- Does the server's slot support the card's full speed?
- Is the card set up for InfiniBand or Ethernet, and can it be changed later?
- Which switches, cables and transceivers has it been validated with?
- If we stay on our existing switches for now, what speed will we actually get?
Parts on SourceIT for these setups
| You need | Listed on SourceIT |
|---|---|
| 🖥️ A desktop AI system with ConnectX-7 | NVIDIA DGX Spark, ASUS Ascent GX10, Lenovo ThinkStation PGX |
| 🔌 A cable to stack two of them | 200G QSFP56 DAC cable for GB10 ConnectX-7 stacking, 1 m / 2 m / 3 m |
| 🔀 A 200 GbE switch for three or more | NVIDIA Spectrum-2 SN3700, 32-port 200GbE |
| 🔀 A compact 400G switch | MikroTik CRS804 DDQ, 4-port 400G QSFP56-DD |
| 🔌 A short 400G cable inside the rack | NVIDIA QSFP112 400G DAC cable, 40 cm |
| 🧠 A server with ConnectX-8 built in | HPE Compute XD690 and the other HGX B300 servers listed above |
Listings and availability change. Check each product page for current details, and ask us to confirm compatibility for your exact setup.




