Can You Cluster Different GB10 Systems? Mixing DGX Spark, ASUS, Dell, HP, MSI, Lenovo, Gigabyte and Acer

You own a DGX Spark and you want to add an ASUS Ascent GX10 because it was in stock. Or you have a Dell Pro Max GB10 from work and an HP ZGX Nano you bought yourself. Can they be clustered together, or do all the boxes have to come from the same vendor? They can be clustered together. All eight GB10 systems on the market are built on the same NVIDIA GB10 Grace Blackwell reference design with the same ConnectX-7 200GbE NIC and the same two QSFP ports. To the cable, to NCCL and to the network, a DGX Spark and an ASUS GX10 are the same machine. Owners are running mixed clusters today, from two-node pairs up to an eight-node build with four brands in it.

What does differ between brands is software, firmware and support — not cabling. Below: which mixes have been run in the field, what to buy at each node count, the four things that actually trip up mixed clusters, and a setup checklist.

Why Different GB10 Brands Work Together

The GB10 systems are not eight different computers that happen to use the same chip. They are eight chassis around one NVIDIA reference board. Every one of them has the GB10 Grace Blackwell Superchip, 128 GB of unified LPDDR5x memory, and a ConnectX-7 NIC with two 200GbE QSFP ports wired to the SoC the same way. The vendors differ in cooling, power adapter, front-panel details, SSD size and support contract — none of which the cluster network sees.

That matters because clustering on GB10 happens at three layers, and the brand logo touches none of them:

• The cable sees a ConnectX-7 QSFP cage. Same cage on every brand.
• The network sees RoCE over Ethernet between ConnectX-7 NICs. Same NIC, same firmware family.
• NCCL sees GB10 GPUs and ConnectX-7 interfaces. Same GPU, same interfaces, same four-logical-interfaces-for-two-ports layout.

Which Mixes Have Actually Been Run

This is not just theory. Owners on the NVIDIA developer forums report running mixed clusters with no networking problems, and ServeTheHome built an eight-node GB10 cluster that combines NVIDIA, ASUS, Dell and Lenovo units on a 400GbE switch. Here is where each system stands as of October 2026:

One thing no vendor does is document mixed-brand clusters. NVIDIA's clustering guide is written for DGX Spark, and each partner documents clustering its own units. A mixed cluster is proven in the field, not certified by anyone — which mostly matters when you open a support ticket (more on that below).

What to Buy at Each Node Count

Mixing brands does not change the topology rules at all. Each node has two QSFP ports, and the same limits apply whatever logo is on the chassis:

For the details behind the four-node rows — why breakout cables on the GB10 side do not work, how to fold a ring in a rack, and when the switch is the better buy — see Can You Build a 4-Node DGX Spark Cluster Without a Switch? For the switch build itself, see MikroTik CRS804 + DGX Spark: the exact cables for a 4-node cluster.

The Four Things That Actually Differ Between Brands

1. The OS image

Most GB10 systems ship with NVIDIA DGX OS, the Ubuntu-based image NVIDIA builds for the platform, but each vendor ships its own build of it and some SKUs are sold with a vendor Ubuntu image instead. A cluster wants the same OS version, the same CUDA and driver versions and the same NCCL on every node. Owners on the NVIDIA forums report the cleanest route is to install NVIDIA's DGX Spark recovery image on the partner box as well, after which the two systems run an identical OS and receive updates together. One owner did this on an ASUS GX10; another reimaged an HP ZGX Nano with NVIDIA's image after an SSD upgrade with no issues.

2. Firmware lag

Partner firmware does not always land the same day as NVIDIA's. One HP ZGX Nano owner reported their unit's firmware running about two months behind the DGX Spark it was clustered with. Firmware includes the ConnectX-7 and the SoC, so check versions on every node before you chase a throughput problem — a firmware mismatch looks exactly like a cabling or RoCE problem from the outside.

3. Port location and labeling

The ports are electrically the same, but where they sit on the rear panel, and what the vendor calls them, differ by chassis. Some spec sheets call them QSFP, some ConnectX-7, and at least one labels them in InfiniBand terms. GB10 clustering runs over Ethernet (RoCE), so whatever the label says, confirm every port reports an Ethernet link layer — ibv_devinfo shows link_layer: Ethernet — before you build the NCCL configuration.

4. Support

Each vendor supports its own box. If you reimage a partner system with NVIDIA's image, or open a ticket about a cluster that includes another brand, expect the conversation to start with "does it happen on a single unit?" Check your vendor's warranty terms before reimaging, and keep a single-node reproduction handy for any issue you report.

Will Mixing Brands Slow the Cluster Down?

Not on the network: every brand has the same ConnectX-7 and the same roughly 200 Gb/s per node, so links between brands run at the same speed as links within one. The places a mixed cluster can be uneven are outside the network:

• Cooling. Chassis designs differ, and tensor-parallel work moves at the pace of the slowest node. If one brand runs hotter under sustained load, the whole cluster waits for it. Give every unit the same airflow and watch clocks under a long run.
• Storage. SSDs range from 1 TB to 4 TB across SKUs. Each node typically needs its own copy of the model weights, so the smallest drive in the cluster sets how many large models you can keep locally. A 4 TB head node with 1 TB workers is a common and workable setup — just plan model storage around the 1 TB.
• Memory. All eight systems have 128 GB, so there is no memory mismatch to balance.

Mixed GB10 Cluster Setup Checklist

• Same OS image and version on every node — reimage with NVIDIA's DGX Spark image if the partner image differs (check warranty terms first)
• Same firmware, CUDA, driver and NCCL versions on every node
• Confirm each port's rear-panel location in that vendor's manual before cabling
• Confirm every ConnectX-7 port reports link_layer: Ethernet
• Two nodes: one cable. Three nodes: a three-cable ring. Four nodes: a switch, or a four-cable ring with custom NCCL (unsupported)
• Never put a breakout cable on the GB10 side — breakout belongs on the switch side only
• MTU 9000 on every node and switch port; bind both 100G halves of each port
• Plan model storage around the smallest SSD in the cluster

FAQ

Can I cluster a DGX Spark with an ASUS Ascent GX10?
Yes. The DGX Spark and the ASUS Ascent GX10 use the same NVIDIA GB10 board and the same ConnectX-7 200GbE QSFP ports, so they connect with the same cable and run the same NCCL setup as two Sparks. Owners are running Spark + GX10 pairs today. Match the OS and firmware on both before benchmarking.

Can different GB10 brands be clustered together?
Yes. DGX Spark, ASUS Ascent GX10, Dell Pro Max GB10, HP ZGX Nano, MSI EdgeXpert, Lenovo ThinkStation PGX, Gigabyte AI TOP ATOM and Acer Veriton GN100 all share the GB10 reference design and ConnectX-7 ports. Mixed clusters of NVIDIA, ASUS, Dell, HP, Lenovo and Gigabyte units have been run in the field, including an eight-node cluster with four brands. No vendor formally documents mixed-brand clusters, so each vendor supports only its own units.

How many mixed GB10 systems can I connect without a switch?
The same as an all-NVIDIA cluster. Two nodes take one cable and three nodes form a ring with three cables, both documented by NVIDIA. Four nodes can run as a ring with four cables using a custom NCCL build, which works but is not NVIDIA-supported. The supported route for four nodes, and the only route past four, is a switch.

Do all nodes need the same operating system?
They should run the same OS version, CUDA, driver and NCCL. Most GB10 systems ship with NVIDIA DGX OS, but vendor builds and update timing differ. Owners report the simplest fix is to install NVIDIA's DGX Spark image on the partner units so every node runs an identical OS — check your vendor's warranty terms before reimaging.

How much does the GB10 cluster cable cost?
The NJAAKK-N911 compatible QSFP112 400G stacking cable, 0.5 m, is $174 at Resilient Tec, and the MCP1650-V00AE30 200G QSFP56 DAC, 0.5 m, is $84 at Resilient Tec. Both are new with a lifetime warranty. A three-node ring is $522 in cabling and a four-node ring is $696. (Prices checked October 12, 2026.)

Where can I buy a cable to cluster different GB10 systems?
Directly from Resilient Tec, shipping from Knoxville, Tennessee, with POs and net terms available. The NJAAKK-N911 compatible stacking cable is $174 at resilient-tec.com and the MCP1650-V00AE30 200G DAC is $84 at resilient-tec.com, both new with a lifetime warranty. (Prices checked October 12, 2026.)

Does mixing GB10 brands reduce performance?
Not on the network — every brand has the same ConnectX-7 and roughly 200 Gb/s per node. Differences show up elsewhere: chassis cooling can make one brand hold clocks differently under sustained load, and tensor-parallel work runs at the pace of the slowest node; and the smallest SSD in the cluster limits how many model copies you can keep locally.

Ready to Buy?

The same parts link any GB10 brand to any other. New, lifetime warranty, shipping from Knoxville, Tennessee. Prices below checked October 12, 2026.

• NJAAKK-N911 compatible — QSFP112 400G stacking cable, 0.5 m — $174, one for two nodes, three for a ring, four for a switchless four-node ring
• LMTQF022-SD-R compatible — QSFP112 400G stacking cable, 0.5 m — $174, the same cable under Luxshare's part number
• MCP1650-V00AE30 — 200G QSFP56 DAC, 0.5 m — $84, the lower-cost two-node link
• MCP1650-H002E26 — 200G QSFP56 DAC, 2 m — $91, one per node when the switch has a 400G port per node
• MCP7H60-W001 — 400G QSFP-DD to 2×200G QSFP56 breakout, 1 m — $149, switch side, one per pair of nodes
• MCP7H60-W002 — 400G QSFP-DD to 2×200G QSFP56 breakout, 2 m — $179.45, switch side, one per pair of nodes
• MCP7H60-W003 — 400G QSFP-DD to 2×200G QSFP56 breakout, 3 m — $199, when the switch is a few rack units away

Building a mixed cluster at four nodes or more? Send us your models, node count and switch and we will confirm exactly what to buy before you order.

You Might Also Need

• Does the DGX Spark cable work with every GB10 system? — the per-brand cable answer for ASUS, Dell, HP, MSI, Lenovo, Gigabyte and Acer
• Can you build a 4-node DGX Spark cluster without a switch? — the switchless ring, the breakout trick that fails, and when to buy a switch
• MikroTik CRS804 + DGX Spark — the exact cables for a four-node switch build
• GB10 stuck at 100 Gb/s on a breakout? — the fixes that get you to 200G
• DGX Spark ports explained — why USB-C isn't your clustering cable
• DGX Spark interconnect reference — part numbers and what to buy at two, three and four or more nodes

Sources

• NVIDIA DGX Spark User Guide — clustering
• NVIDIA Developer Forums — Mixing different GB10 systems in a DGX Spark cluster
• NVIDIA Developer Forums — How many DGX Spark/GB10 devices do you have?
• ServeTheHome — BIG AI Cluster Little Power: the 8x NVIDIA GB10 Cluster
• Tom's Hardware — Local AI clustering with Dell's Pro Max GB10