Short answer: in Resilient Tec customer James Lee's four-node GB10 cluster, a direct ring and a switch performed about the same. NCCL all-reduce bus bandwidth was 23.8 GB/s on the ring and 24.1 GB/s through a MikroTik CRS812 switch. Qwen3-235B total output throughput was within about 3% at 4 to 64 concurrent requests, with the ring about 10% ahead on a single request. The switch doesn't make the cluster faster. It gives you NVIDIA's supported setup and room to add more nodes.
These are progress results on stock or partly tuned settings, not final numbers. We'll update this post when the follow-up runs are done.
Short answer: same speed, the switch buys room to grow
Four DGX Sparks, ring or switch, on $79 MCP1650-V00AE30 cables at Resilient Tec
Both builds used Resilient Tec's MCP1650-V 200G QSFP56 passive DACs, straight, one per link, with no breakouts. A four-node ring takes four 0.5 m cables, $316 in all. A switch build takes one cable from each Spark to the switch, usually 1–2 m.
|
MCP1650-V00AE30 — 200G QSFP56 passive DAC, 0.5 m Node to node in a ring. Four for a four-node ring: $316. |
$79 · In stock at Resilient Tec Add to cart |
|
MCP1650-V001E30 — 200G QSFP56 passive DAC, 1 m Spark to switch when the switch sits beside the stack. Four: $388. |
$97 · In stock at Resilient Tec Add to cart |
|
MCP1650-V002E26 — 200G QSFP56 passive DAC, 2 m Spark to switch a few rack units away. Four: $460. |
$115 · In stock at Resilient Tec Add to cart |
Resilient Tec sells the MCP1650-V00AE30 200G QSFP56 DAC for $79 directly at resilient-tec.com, with the 1 m at $97 and the 2 m at $115 — new, lifetime warranty, shipping from Knoxville, Tennessee. Prices checked November 5, 2026.
The setup
- Nodes: four Acer Veriton GN100 AI Mini Workstations. Each is a GB10 system, the same chip, ConnectX-7 NIC and two QSFP ports as the NVIDIA DGX Spark.
- Cables: Resilient Tec MCP1650-V 200G QSFP56 passive DACs (compatible with NVIDIA's MCP1650-V), straight, no breakouts
- Topology A: switchless ring, four straight cables, each node wired to its two neighbors
- Topology B: MikroTik CRS812 switch, one straight MCP1650 from each node. Two went into the switch's 200G QSFP56 ports and two into its 400G QSFP56-DD ports.
All four switch links came up at 200G, and transfer tests measured about 196 Gb/s per link.
"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, the same ones listed in Ready to Buy below.
Results
| Test | Ring | Switch | Difference |
|---|---|---|---|
| NCCL all_reduce_perf, 8 GiB, bus bandwidth | 23.8 GB/s | 24.1 GB/s | Switch +1.3% |
| Qwen3-235B-A22B GPTQ-Int4, 1 concurrent request | 33.4 tok/s | 30.2 tok/s | Ring +10.6% |
| Same, 4 concurrent | 75.3 tok/s | 73.3 tok/s | Ring +2.7% |
| Same, 16 concurrent | 143.9 tok/s | 139.6 tok/s | Ring +3.1% |
| Same, 64 concurrent | 254.2 tok/s | 255.4 tok/s | Switch +0.5% |
| DeepSeek, best real-text run so far (transport-tuned) | 78.23 tok/s | 82.66 tok/s | Not comparable, see note |
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 don't credit that gap to the cables or the switch alone. 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.
What this means
- The ring and the switch are close. The ring was ahead with a single request, by about 10%, and within about 3% at 4 and 16 requests. The switch was slightly ahead at 64 requests and on NCCL bandwidth.
- Straight MCP1650 cables work in the CRS812 at 200G, in both its 200G QSFP56 ports and its 400G QSFP56-DD ports.
- Breakout cables aren't required to connect four Sparks to a CRS812. Their benefit is fewer switch ports used, which leaves room to expand.
- The switch is the supported way to build a four-node cluster and the only way to grow past four. The switchless ring works too, but owners run it with custom NCCL configuration, and NVIDIA doesn't support it. For the full topology rundown, including why a breakout mesh on the Spark side doesn't work, see Can You Build a 4-Node DGX Spark Cluster Without a Switch?
- Staying at three nodes? You don't need either. Three Sparks connect directly in an NVIDIA-documented ring: see How to Connect Three DGX Sparks.
What each build costs in cables
| Build | Cables | Cable total | Also needed |
|---|---|---|---|
| Switchless ring | 4 × MCP1650-V00AE30, 0.5 m | $316 | None. Unsupported, custom NCCL, stops at four |
| Switch, units beside it | 4 × MCP1650-V001E30, 1 m | $388 | A 400G-class switch such as the CRS812 |
| Switch, a few rack units away | 4 × MCP1650-V002E26, 2 m | $460 | A 400G-class switch such as the CRS812 |
Prices checked November 5, 2026. Switch pricing and the RoCE tuning that goes with it are covered in MikroTik CRS804 + DGX Spark.
What's next
James will rerun the same tests with the same software settings once breakout cables arrive, and he's checking answer quality alongside speed. We'll publish that as part two. The breakout he's testing is the switch-side MCP7H60-W001 400G QSFP-DD to 2×200G QSFP56 type, $149 at Resilient Tec.
Ready to Buy?
The cables behind both builds, the same Resilient Tec MCP1650-V DACs James tested. New, lifetime warranty, shipping from Knoxville, Tennessee. Prices below checked November 5, 2026.
- MCP1650-V00AE30 — 200G QSFP56 passive DAC, 0.5 m — $79, in stock at Resilient Tec; four for a four-node ring ($316)
- MCP1650-V001E30 — 200G QSFP56 passive DAC, 1 m — $97, in stock at Resilient Tec; Spark to switch, four for a four-node switch build ($388)
- MCP1650-V002E26 — 200G QSFP56 passive DAC, 2 m — $115, in stock at Resilient Tec; Spark to switch a few rack units away ($460 for four)
- MCP1650-V003E26 — 200G QSFP56 passive DAC, 3 m — $145, in stock at Resilient Tec; for longer switch runs
- MCP7H60-W001 — 400G QSFP-DD to 2×200G QSFP56 breakout, 1 m — $149, in stock at Resilient Tec; switch side only, one cable feeds two Sparks
Need a different length or a quote for a full cluster? Send us your node count and switch model and we'll confirm exactly what to buy. POs and net terms available.
FAQ
Is a switch faster than a direct ring for 4 DGX Sparks?
Not in these tests. NCCL all-reduce bus bandwidth was 24.1 GB/s through a MikroTik CRS812 switch and 23.8 GB/s on the switchless ring. Qwen3-235B total throughput was within about 3% at 4, 16 and 64 concurrent requests. The ring was about 10% faster with a single request. The switch's advantage is room to add nodes and NVIDIA support, not speed.
How much do cables for a 4-node DGX Spark cluster cost?
A switchless four-node ring takes four MCP1650-V00AE30 200G QSFP56 DACs at $79 each from Resilient Tec, $316 in all. A switch build takes one cable per node from the Spark to the switch: four 1 m MCP1650-V001E30 at $97 ($388) or four 2 m MCP1650-V002E26 at $115 ($460), plus the switch. All are new with a lifetime warranty. (Prices checked November 5, 2026.)
Where can I buy MCP1650 cables for a DGX Spark cluster?
Directly from Resilient Tec at resilient-tec.com, shipping from Knoxville, Tennessee, with POs and net terms available. The MCP1650-V00AE30 0.5 m is $79, the 1 m is $97, the 2 m is $115 and the 3 m is $145. All are Ethernet-coded 200G QSFP56 passive DACs, new, with a lifetime warranty. (Prices checked November 5, 2026.)
Do straight MCP1650 cables work in a MikroTik CRS812?
Yes. James Lee's cluster used Resilient Tec's compatible MCP1650-V cables, and 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?
Not on a CRS812. It 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. On a switch with only two 400G ports and nothing else that takes a 200G link, you need QSFP-DD to 2×200G breakouts on the switch side.
Is a switchless 4-node DGX Spark ring supported by NVIDIA?
No. NVIDIA documents direct cabling for up to three DGX Sparks and a switch for four. A four-node ring works, as these results show, but owners run it with custom NCCL configuration outside NVIDIA support, and it cannot grow past four. See Can You Build a 4-Node DGX Spark Cluster Without a Switch?
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 DGX Spark. These benchmarks ran on four of them, with the same cables and topologies a DGX Spark uses. For mixing brands, see Can You Cluster Different GB10 Systems?
You Might Also Need
- Can you build a 4-node DGX Spark cluster without a switch? — the four-cable ring, the breakout mesh that doesn't work, and when to buy a switch
- How to connect three DGX Sparks — the NVIDIA-documented three-node ring, port by port
- Can you cluster different GB10 systems? — DGX Spark, ASUS, Dell, HP, Acer and other GB10 brands in one cluster
- MCP1650-V00AE30 vs NJAAKK-N911 — the 200G DAC used here vs the QSFP112 stacking cable
- GB10 stuck at 100 Gb/s? — the configuration fixes that get you to 200G
- How to connect two NVIDIA DGX Spark systems — if you're starting with two
- DGX Spark interconnect reference — part numbers and what to buy at two, three and four or more nodes