400Gbps Networking for AI: Do You Really Need Such an Expensive Switch?

For two or three NVIDIA GB10 systems, no switch at all: two connect with one cable and three can be wired in a ring. At four units a switch becomes mandatory, but it does not need to be expensive. NVIDIA's requirement is one 200 Gb/s Ethernet port per device, and a compact switch such as the MikroTik CRS804 DDQ (S$2,255 at SourceIT) can provide eight of those by splitting its four 400G ports. A full data-centre switch such as the NVIDIA Spectrum-2 SN3700 (S$28,523.25) makes sense when the cluster lives in a rack, grows beyond eight units or shares a fabric with storage and servers. The "400G" in the product names is the port size on the switch; each GB10 talks at 200 Gb/s. Prices listed 28 Oct 2026.

Do you need a switch at all?

From NVIDIA's Cluster Assistant documentation, DGX Spark playbooks and NVIDIA's October 2026 announcement on eight-unit configurations.

Bandwidth vs latency: what does AI traffic actually need?

When a model is split across GB10s, the units exchange data constantly: for inference, small messages for every layer of every token; for training, large blocks of gradients. Two properties matter:

• Bandwidth is how much data moves per second. It matters most for training and for loading large models.
• Latency is how long each message takes to arrive. It matters most for inference, because generation waits on many small exchanges per token.

Published results suggest inference is far less bandwidth-hungry than the port numbers imply. In a community eight-unit build reported on NVIDIA's forum in June 2026, running each node at 100G instead of 200G changed per-stream generation speed by only about 3%. Measured link usage is also below the headline: forum users report NCCL all-reduce bus bandwidth of about 27 GB/s on GB10. NVIDIA publishes no latency figures for the GB10's ConnectX-7. The practical reading: for inference, a switch that runs at 200G per port with low latency is plenty; paying for more bandwidth per node buys little.

For inference clusters, the switch needs to be fast enough and simple enough, not the biggest you can afford.

The two switch options compared

From MikroTik and NVIDIA specifications. MikroTik's page does not mention RoCE features such as PFC or ECN; NVIDIA does not name a recommended switch for DGX Spark clusters.

When is the MikroTik enough?

For a four to eight-unit GB10 inference cluster on a shelf or in a small cabinet, it is the obvious starting point: a fraction of the cost of one GB10, low power, and sized for exactly this job. Community users on NVIDIA's forum describe running GB10 clusters on it, with two practical tips: use QSFP56-DD to 2x QSFP56 breakout cables, and set the ports to 200G manually rather than relying on auto-negotiation. Some also report link aggregation problems, and supply has been tight, with users citing backorders into 2027. Ask us for current lead time.

When does the SN3700 earn its price?

• More than eight units, or a plan to grow there.
• Training workloads, where gradient exchange benefits from a switch built for lossless RoCE fabrics.
• A shared fabric connecting GB10s to fast storage or GPU servers.
• A data-centre or server room home, with rack space, cooling and a UPS. NVIDIA's manual lists 250 W and a 0 to 40 °C operating range.
• Standardisation on NVIDIA networking and Cumulus Linux across the organisation.

Ring or switch at three units?

A ring of three direct cables avoids buying a switch and NVIDIA's tools support it. A switch is cleaner if you expect to add a fourth unit, and NVIDIA's own eight-unit tests in October 2026 found switched topologies faster than rings: on Qwen3.8-27B, four units scored 36.3 tokens per second switched against 32.8 in a ring, and eight units 52.6 switched against 40.1 in a ring. NVIDIA also notes its new all-to-all communication backend works only on switched topologies. If growth is likely, buy the switch at three.

Do faster cables help?

No. NVIDIA's user guide is explicit: "Faster cables don't help because the port is capped at 200 Gb/s." Each GB10 QSFP port is Ethernet-only and runs at up to 200 Gb/s, even with a 400G-rated cable. Buy cables on NVIDIA's approved list, or ones matched to your switch, not the fastest available. The link itself is explained in NVIDIA ConnectX-7 vs Ordinary Ethernet: Why AI Clusters Need More Than Fast Internet, and the wider question of who needs this class of networking in 400Gbps Networking on a Desk: Who Actually Needs This Monster?

A buying checklist

1. Two units: buy one approved QSFP112 cable, such as the NVIDIA QSFP112 400G DAC (S$155). No switch.
2. Three units: three cables in a ring, or a switch if a fourth is likely.
3. Four to eight units: the CRS804 DDQ with breakout cables, ports set to 200G.
4. Beyond eight, training fabrics or rack deployments: the SN3700.
5. Put the switch on the same UPS as the units.
6. Run NVIDIA's link speed test after cabling; each link should approach 200 Gb/s.

For whether a cluster is worth building, see Four GB10 Supercomputers on One Desk and Can Eight NVIDIA GB10 Systems Compete With Enterprise AI Infrastructure?

Buying in Singapore

SourceIT supplies the MikroTik CRS804 DDQ, the NVIDIA Spectrum-2 SN3700, QSFP cables and GB10 systems from ASUS, NVIDIA, HP, Dell, Lenovo, GIGABYTE and MSI, with local warranty and a GST invoice, and can specify matching cables and commission the cluster. Browse the NVIDIA DGX Spark collection. 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

Do I need a switch to cluster DGX Sparks? Not for two units, which connect directly, or three, which can be wired in a ring. Four units require a switch with a 200 Gb/s Ethernet port for each device.

Do I need a 400G switch for GB10? You need 200 Gb/s per unit. Switches with 400G ports, such as the MikroTik CRS804 DDQ, are useful because each port splits into two 200G links.

Does a faster cable make a GB10 cluster faster? No. NVIDIA states the port is capped at 200 Gb/s, so faster cables do not help.

Which switch does NVIDIA recommend? NVIDIA names no specific model. Its requirement is one 200 Gb/s Ethernet link per device; the MikroTik CRS804 DDQ and NVIDIA SN3700 are both options.

Is a ring or a switch faster? In NVIDIA's October 2026 tests, switched topologies were faster: four units reached 36.3 tokens per second switched against 32.8 in a ring on Qwen3.8-27B.