DGX Spark 4-Node Cluster Benchmark Data

Short answer: in Resilient Tec customer James Lee's four-node GB10 cluster, a switchless ring and a MikroTik CRS812 switch performed about the same. NCCL all-reduce bus bandwidth was 23.84 GB/s on the ring, 24.12 GB/s through the switch with four straight cables, and 24.10 GB/s with two switch-side breakout cables. The switch does not make a four-node cluster clearly faster. It is the NVIDIA-supported setup, and it leaves room to add nodes.

These are progress results on stock or partly tuned settings, not final numbers. The narrative write-up, with photos and buying notes, is 4x DGX Spark: Ring vs. Switch, Real Benchmarks. This page is the citable data: summary findings, results tables, topologies, methodology, and how to credit the work.

Data collected by Resilient Tec customer James Lee. Photos courtesy of James Lee.

Summary findings

• The ring stayed within about 1% of the switch on NCCL all-reduce (23.84 GB/s vs 24.12 GB/s with four individual cables, and 24.10 GB/s with two breakouts).
• On NCCL all-gather, the switch setups were slightly ahead (23.79 and 23.80 GB/s vs 23.25 GB/s on the ring). On send/receive, the ring was ahead (23.96 GB/s vs 22.89 and 22.88 GB/s), about 4.5%.
• Qwen3-235B-A22B GPTQ-Int4 total output throughput was within about 3% at 4, 16 and 64 concurrent requests. The ring was about 10% ahead on a single request (33.4 vs 30.2 tok/s).
• Two 400G-to-2×200G breakout cables on the switch side performed essentially the same as four individual cables on NCCL and on raw transfers (196.07 to 196.08 Gb/s per link, including with all four links active, compared with about 196.1 Gb/s on individual cables).
• Later model-speed runs on DeepSeek V4.1 Flash and Qwen3.8 Flash Next also show the breakouts are not slowing things down. They are not a cable-only comparison: the breakout figures are averages of warmed-up runs after a shop move and later switch tuning, the ring and individual-cable figures are earlier single runs, and the Qwen ring run used different settings. Do not read those rows as proof that breakouts make models faster.
• A four-node switchless ring with four straight cables works, but it needs custom NCCL configuration and is not NVIDIA-supported. It stops at four. A switch is the supported route. NVIDIA has officially supported eight-node clusters since October 22, 2026; NVIDIA Sync still sets up at most four through a switch.
• A breakout mesh on the Spark side does not work. Switch-side breakouts do.

Topologies

All three setups used four Acer Veriton GN100 AI Mini Workstations. Each is a GB10 system, with the same chip, ConnectX-7 NIC and two QSFP ports as the NVIDIA DGX Spark.

• A. Switchless ring. Four straight 200G QSFP56 DACs, each node wired to its two neighbors. No switch.
• B. Switch, four individual cables. A MikroTik CRS812, one straight MCP1650-V cable from each node. Two cables went into the switch's 200G QSFP56 ports and two into its 400G QSFP56-DD ports. All four links came up at 200G.
• C. Switch, two breakouts. The four individual cables to the CRS812 were replaced with two 400G-to-2×200G breakout cables. The 400G end plugs into the switch; each 200G end goes to one Spark. All four Sparks still get their own 200G connection, and two switch ports are left free for a fifth and sixth Spark.

"The straight NVIDIA MCP1650 cables are working with the MikroTik CRS812. All four Sparks are connected, including two through the 400G ports. All four links negotiated at 200G and transfer tests measured about 196 Gb/s. So the breakout cables aren't needed to connect four Sparks, but they'll free up ports for expansion."

— James Lee, Resilient Tec customer

The MCP1650 cables James refers to are Resilient Tec's compatible MCP1650-V DACs, listed under Hardware and cables below.

Results

Per-link transfers, part two

Raw transfers on the breakouts measured 196.07 to 196.08 Gb/s per link, including with all four links active at once, compared with about 196.1 Gb/s on individual cables. The error, drop and error-correction counters James checked did not increase. The GPU communication tests ran over RoCE and finished cleanly with zero data errors. A few pause frames appeared, which are flow control, not data errors.

NCCL, 8 GiB, bus bandwidth

Four-GPU NCCL tests. The earlier ring and individual-cable runs were done before James moved the rig from a too-hot closet to his shop and before later switch tuning, so this is not a strict back-to-back cable-only comparison.

23.8–24.1 GB/s of NCCL bus bandwidth is about 190 Gb/s, close to the 200 Gb/s each Spark can move.

Qwen3-235B and DeepSeek, ring vs switch

These ring-versus-switch application results were published with the original post, before the breakout follow-up. Qwen3 numbers are total output tokens per second across all requests, not per request. They used NVIDIA's 26.10 vLLM image, tensor parallel across all four nodes, 0.80 GPU memory utilization and stock settings. The DeepSeek runs include transport tuning, so do not credit that gap to the cables or the switch alone.

Model speed on all three setups

James later ran real models on the two-breakout setup. Here are all three setups side by side, in output tokens per second.

The breakout numbers are averages of warmed-up runs. The ring and individual-cable numbers are earlier single runs, taken before the rig moved out of a too-hot closet and before later switch tuning, and the Qwen ring run used different settings. So this is not a cable-only comparison, and it does not show that breakouts make models faster. It does show the breakouts aren't slowing anything down. James is still checking answer quality alongside speed.

Methodology and caveats

• Who ran it. Resilient Tec customer James Lee, on his own four-node cluster. Resilient Tec supplied compatible MCP1650-V cables and is publishing the results with his credit and photos. These are not a lab certification and not an NVIDIA-sponsored test.
• Nodes. Four Acer Veriton GN100 AI Mini Workstations (GB10, ConnectX-7, two QSFP ports), the same interconnect as a DGX Spark.
• Cables, ring and individual-cable switch tests. Resilient Tec MCP1650-V 200G QSFP56 passive DACs (compatible with NVIDIA's MCP1650-V), straight, one per link, no breakouts.
• Cables, breakout follow-up. Two 400G-to-2×200G breakout cables on the switch side only. The exact length James used was not recorded in the published write-up, so it is not listed as a tested SKU here. The MCP7H60-class cables below are the parts sold for this layout.
• Switch. MikroTik CRS812. In the four-cable setup, two links used the 200G QSFP56 ports and two used the 400G QSFP56-DD ports.
• Link checks. All four switch links negotiated at 200G. Transfer tests measured about 196 Gb/s per link on individual cables, and 196.07 to 196.08 Gb/s on the breakouts, including with all four links active. Error, drop and error-correction counters did not increase on the breakouts. RoCE tests finished with zero data errors; pause frames are flow control.
• NCCL. Four-GPU tests, 8 GiB, bus bandwidth, including all_reduce_perf, all-gather and send/receive.
• Qwen3-235B-A22B GPTQ-Int4. NVIDIA's 26.10 vLLM image, tensor parallel across all four nodes, 0.80 GPU memory utilization, stock settings. Throughput is total output tokens per second across all requests.
• DeepSeek (original ring vs switch row). Includes transport tuning, so that gap is not a cable-only or switch-only result.
• Later model rows (DeepSeek V4.1 Flash and Qwen3.8 Flash Next). Breakout figures are averages of warmed-up runs after the shop move and later switch tuning. Ring and individual-cable figures are earlier single runs. The Qwen ring run used different settings.
• Change of location. The earlier ring and individual-cable runs were taken in a too-hot closet. The cluster was later moved to James's shop, with later switch tuning. Treat cross-era rows as directional, not as a controlled A/B of cables alone.
• Progress results. Stock or partly tuned settings, not final numbers. Answer quality is still being checked alongside speed.
• Support status. The four-node ring works and is reported here because it was measured. It is not NVIDIA-supported. Do not read these results as a recommendation to skip a switch if you want a supported setup or plan to grow past four.

NVIDIA's October 22, 2026 announcement of official support for eight-node DGX Spark clusters is here: Running Frontier-Class AI Models Across 8 NVIDIA DGX Spark Nodes.

Hardware and cables

Prices and availability below were checked on November 10, 2026 from the live product pages. The 0.5 m, 1 m and 2 m MCP1650-V totals match the cable math in the benchmark post.

If you need a different length or a quote for a full cluster, send your node count and switch model.

How to cite this data

Please credit James Lee as the person who collected the measurements, and Resilient Tec as the publisher.

Lee, James. DGX Spark 4-node cluster benchmark data: switchless ring vs MikroTik CRS812 vs switch-side breakouts. Resilient Tec, 5 Nov. 2026, updated 9 Nov. 2026, https://resilient-tec.com/pages/dgx-spark-4-node-cluster-benchmark-data.

The narrative write-up is 4x DGX Spark: Ring vs. Switch, Real Benchmarks.

Related reading

• Can you build a 4-node DGX Spark cluster without a switch? — the four-cable ring, the breakout mesh that does not work, and when to buy a switch
• How to connect three DGX Sparks — the NVIDIA-documented three-node ring
• Can you cluster different GB10 systems? — DGX Spark, ASUS, Dell, HP, Acer and other GB10 brands
• MikroTik CRS804 + DGX Spark — port math and the RoCE notes that go with a switch build
• DGX Spark interconnect reference — part numbers and what to buy at two, three and four or more nodes

FAQ

Is a switch faster than a direct ring for 4 DGX Sparks?
Not in these tests. NCCL all-reduce bus bandwidth was 24.12 GB/s through a MikroTik CRS812 with four individual cables and 23.84 GB/s on the switchless ring, a gap of about 1%. Qwen3-235B total output throughput was within about 3% at 4, 16 and 64 concurrent requests. The ring was about 10% ahead with a single request. The switch's advantage is NVIDIA support and room to add nodes, not a clear speed win at four nodes.

Do breakout cables slow down models on a 4-node DGX Spark cluster?
Not in this customer's tests. On two 400G-to-2×200G breakouts through a MikroTik CRS812, DeepSeek V4.1 Flash ran at 85.15 tokens/sec and Qwen3.8 Flash Next at 58.26, compared with 82.66 and 54.09 on four individual cables and 78.23 and 50.42 on a switchless ring. The breakout numbers are averages of warmed-up runs taken after a shop move and later switch tuning, and the Qwen ring run used different settings. This shows breakouts do not slow things down, not that they make models faster. NCCL all-reduce was 24.10 GB/s on the breakouts and 24.12 GB/s on four individual cables.

Is a switchless 4-node DGX Spark ring supported by NVIDIA?
No. NVIDIA documents a direct ring for up to three DGX Sparks. A four-node switchless ring with four straight cables works, as these results show, but owners run it with custom NCCL configuration, and NVIDIA does not support it. A switch is the supported way to build four nodes and the way to grow past four. NVIDIA has officially supported eight-node clusters since October 22, 2026. NVIDIA Sync still sets up at most four devices through a switch.

Do straight MCP1650 cables work in a MikroTik CRS812?
Yes. James Lee's cluster used Resilient Tec's compatible MCP1650-V 200G QSFP56 passive DACs. All four links negotiated at 200G, including the two in the CRS812's 400G QSFP56-DD ports, and transfer tests measured about 196 Gb/s per link.

Do I need breakout cables for 4 Sparks on a CRS812?
No. A CRS812 has two 200G QSFP56 ports and two 400G QSFP56-DD ports, so four straight cables land four Sparks. Breakouts are optional there and mainly save switch ports for growth. In this test, two switch-side 400G-to-2×200G breakouts matched individual-cable NCCL results and left two switch ports free. A breakout mesh on the Spark side does not work.

Does an Acer Veriton GN100 cluster like a DGX Spark?
Yes. The Acer Veriton GN100 is a GB10 system with the same ConnectX-7 NIC and two QSFP ports as the NVIDIA DGX Spark. These benchmarks ran on four of them, with the same cables and topologies a DGX Spark uses.

How should I cite this data?
Lee, James. DGX Spark 4-node cluster benchmark data: switchless ring vs MikroTik CRS812 vs switch-side breakouts. Resilient Tec, 5 Nov. 2026, updated 9 Nov. 2026, https://resilient-tec.com/pages/dgx-spark-4-node-cluster-benchmark-data. The narrative write-up is at https://resilient-tec.com/blogs/news/4x-dgx-spark-ring-vs-switch-real-benchmarks.