The NVIDIA ConnectX-8 SuperNIC is what lets a DGX Station GB300 grow beyond one box, but on its own it does nothing until you build the network around it. Every DGX Station GB300 has two QSFP112 ports at 400 Gb/s each, up to 800 Gb/s in total, plus a 10 GbE RJ45 port for the office network and a 1 GbE port for the management controller. Two stations can be linked directly, with no switch, into what NVIDIA describes as "over 1.4 TB of coherent memory and up to 40 petaFLOPS". The ConnectX-8 also hosts two of the four M.2 drive slots, giving the GPU a fast path to local data. Using the 400G ports needs cables or optics, and for more than two systems a 400G-class switch. Most office networks have neither.
What networking does a DGX Station have?
| Port | Speed | What it is for | What you need to use it |
|---|---|---|---|
| 2× QSFP112 (ConnectX-8) | 400 Gb/s each, up to 800 Gb/s | Linking DGX Stations, AI cluster fabrics, NVMe-over-fabric storage | QSFP112 direct-attach cables or optics (not included); a 400G-class switch beyond two systems |
| 1× RJ45 | 10 Gb/s | Office network, user access, NAS, updates | A 10 GbE switch port and Cat6A cabling |
| 1× RJ45 (BMC) | 1 Gb/s | Out-of-band management: power, sensors, remote console | A separate management network |
| 2× M.2 Gen6 (behind ConnectX-8) | PCIe 6.0 x4 per slot | Data and checkpoint drives with a direct path to the GPU | Drives; GPUDirect Storage packages on the OS |
Sources: NVIDIA DGX Station product page and datasheet (Aug 2026); NVIDIA DGX Station development guide (updated 30 Oct 2026); NVIDIA ConnectX-8 hardware documentation (updated 30 Oct 2026); SourceIT's five DGX Station GB300 listings (4 Nov 2026). We have not tested DGX Station networking ourselves.
Why is the ConnectX-8 more than just a fast network card?
Three reasons. First, speed: 400 Gb/s per port is 50 GB/s, against 1.25 GB/s for 10 GbE, so moving model weights, datasets and checkpoints between systems takes seconds rather than minutes. Second, it supports RDMA over Converged Ethernet (RoCE) and GPUDirect RDMA, so data can move between the GPU and a peer device without passing through the CPU. Third, the two PCIe Gen6 M.2 slots on every DGX Station hang off the ConnectX-8, not the Grace CPU. ServeTheHome estimates about 30 GB/s of theoretical bandwidth per Gen6 slot, and StorageReview measured up to 13.1 GiB/s read and 12.0 GiB/s write with NVIDIA's GDSIO GPUDirect Storage test on the MSI WS300.
NVIDIA's ConnectX-8 documentation lists both InfiniBand and Ethernet at up to 800 Gb/s for the chip. The DGX Station materials and OEM listings describe the ports as Ethernet. We found no confirmation that the Station's ports run in InfiniBand mode, so plan on Ethernet with RoCE.
The GPU makes a DGX Station fast. The ConnectX-8 decides whether it stays a single machine or becomes the first node of something bigger.
How much time does faster networking actually save?
The table below shows the theoretical time to move DeepSeek-R1's 689 GB model folder at each line rate. Real transfers are slower, because protocol overhead and the speed of the drives at each end set the true limit, so treat these as best cases.
- 1 GbE (0.125 GB/s): about 92 minutes
- 10 GbE, the Station's RJ45 port (1.25 GB/s): about 9.2 minutes
- 400 GbE, one QSFP112 port (50 GB/s): about 14 seconds
- 800 Gb/s, both ports (100 GB/s): about 7 seconds
Model size from the DeepSeek-R1 repository on Hugging Face (accessed 4 Nov 2026); times calculated from line rate only.
For a single Station loading a model once a week from a NAS, 10 GbE is usually enough. The 400G ports matter when two systems split one model, when training writes large checkpoints to shared storage, or when several users pull different large models through the day.
Can you link two DGX Stations without a switch?
Yes, and NVIDIA documents how. Its "Connect Two Stations" playbook links QSFP0 to QSFP0 and QSFP1 to QSFP1 with two 400G-capable QSFP cables, configures RoCEv2 with an MTU of 9000 on private addresses, and validates the link with ib_write_bw before running NCCL, Ray or vLLM across both nodes. NVIDIA rates it at about 60 minutes and "Medium" risk. Two points to keep in mind:
- There is no NVLink between the two systems. NVIDIA's playbook states that cross-node traffic uses ConnectX-8 RoCEv2. Each Station's 748 GB is coherent inside the box; across two boxes, software such as vLLM or NCCL splits the work over the network.
- Cables are extra. Every SourceIT GB300 listing says QSFP112 transceivers or direct-attach cables are ordered separately; ASUS's listing adds that transceivers are not included unless you choose its -T SKUs. For two systems side by side, a short DAC such as the HP ZGX 0.4m QSFP112 Ethernet DAC (S$365.00, listed 4 Nov 2026) is the simplest option. NVIDIA's recipe uses two, so S$730.00 for the pair. Confirm compatibility with your specific Station before ordering.
Whether two systems actually run faster depends on the workload. Our post on two GB10 systems explains why linking rarely doubles throughput.
What do you need for three or more systems, or fast shared storage?
A switch with 400G-class ports, plus the right optics or cables for each port. ServeTheHome connected three DGX Stations through an 800G switch in its testing. SourceIT stocks the MikroTik CRS804 DDQ 4-port 400G switch (S$2,255.00) and the NVIDIA Spectrum-2 SN3700 32-port 200GbE switch (S$28,523.25), both listed 4 Nov 2026. The CRS804 uses QSFP56-DD ports and the SN3700 uses 200GbE QSFP56 ports, neither of which is the Station's QSFP112. So the cable or optic for each link, and the speed it will actually run at, needs checking. NVIDIA has not published a list of validated switches or optics for the DGX Station that we could find. Ask us to check the full path before you buy. Our posts 400Gbps Networking for AI: Do You Really Need Such an Expensive Switch? and ConnectX-7 vs Ordinary Ethernet cover the switch decision in more depth.
The same fabric can reach fast shared storage. NVIDIA documents GPUDirect Storage with local NVMe, NVMe-over-fabric and several parallel and NFS-class file systems, though NVIDIA's DGX OS guide notes the Station needs extra packages and a driver rebuild to enable it.
What about the ordinary office network?
Most teams will use the 10 GbE RJ45 port for everyday access: SSH, notebooks, model endpoints, updates and pulling data from a NAS. That needs a 10 GbE switch port, such as on the D-Link DXS-1210-10TS (S$1,680.00) or the HPE Networking Instant On 1960 12-port 10G (S$2,444.40). If datasets live on a NAS, give the NAS a 10 GbE link too, for example with a Synology E10G30-T2 dual-port 10GbE card (S$387.60), all listed 4 Nov 2026. Put the 1 GbE management port on a separate network, and keep it away from the internet; see Should Your Company's Most Powerful Computer Be Connected to the Internet?
Networking is also part of resilience. A second Station on the same fabric can take over work if the first fails, which is one answer to the risk in Your DGX Station Could Be a Single Point of Failure. For storage sizing on the box itself, see A S$150,000 AI Workstation With Just a Few Terabytes of SSD Storage?
Buying in Singapore
SourceIT lists all five DGX Station GB300 systems (HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC), price on request, in the NVIDIA DGX Spark & DGX Station collection, along with the switches, cables and NAS hardware above. Tell us whether you plan one station, a linked pair or a switched cluster, and we will quote the cables, optics and switch ports with the systems. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
How fast is the DGX Station's network? Its ConnectX-8 SuperNIC has two QSFP112 ports at 400 Gb/s each, up to 800 Gb/s in total. It also has a 10 GbE RJ45 port for the office network and a 1 GbE management port.
Can two DGX Stations be connected without a switch? Yes. NVIDIA's playbook links two Stations port to port with two 400G-capable QSFP cables, using RoCEv2 and an MTU of 9000. There is no NVLink between the two systems.
Are cables included? No. SourceIT's DGX Station listings state that QSFP112 transceivers or direct-attach cables are ordered separately. The HP ZGX 0.4m QSFP112 DAC is S$365.00 on SourceIT's 4 Nov 2026 listing.
Does the DGX Station support InfiniBand? The ConnectX-8 chip supports InfiniBand and Ethernet, but DGX Station materials describe the ports as Ethernet with RoCE. We found no confirmation of InfiniBand mode on the Station.
Do I need a 400G switch? Only for three or more systems or 400G shared storage. A single Station can use its 10 GbE port, and two Stations can be linked directly.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
