Cluster DGX Spark: Link Two ASUS Ascent GX10 Units (256 GB)
Posted by Wei Fei on
Linking two ASUS Ascent GX10 units is the cheapest route in Singapore to a 256 GB AI machine: one QSFP cable between the ConnectX-7 ports, ten minutes in NVIDIA Sync's Cluster Assistant, and a model that no single GB10 box can hold, such as a Llama 3.1 405B-class checkpoint at FP4, runs across both. This guide covers the hardware, placement, the Cluster Assistant steps and the manual alternative, verifying the 200 GbE link, serving one model across both units with NVIDIA's vLLM playbook, and what changes at three or four units. It is drawn from NVIDIA's DGX Spark documentation and applies equally to the NVIDIA DGX Spark and other GB10 systems.
What clustering GX10s gives you
| Units | Pooled memory | Cabling | Set-up path |
|---|---|---|---|
| 1 | 128 GB | None | Single unit, up to roughly 200B parameters at FP4 |
| 2 | 256 GB | One QSFP cable, direct, same port on each unit | NVIDIA Sync Cluster Assistant or manual netplan |
| 3 | 384 GB | Three cables in a direct ring, or three cables to a switch | Cluster Assistant; NCCL 2.30u1 updated ring-topology support in June 2026 |
| 4 | 512 GB | One cable per unit to a switch with a 200 Gbit/s port for each | Cluster Assistant; direct cabling not supported at four |
Two caveats from NVIDIA's documentation set expectations. The Cluster Assistant configures the network and inter-device SSH only; it does not set up workloads, and Slurm or Kubernetes are yours to add. And "model sharding depends on the model and cluster configuration", which is why NVIDIA's serving playbook limits itself to two units.
Hardware: two GX10s and one cable
You need two GX10s on the same LAN, each reachable by IP or mDNS name, a user with sudo on both, and one direct-attach cable. SourceIT stocks the ASUS QSFP cable for the Ascent GX10 (40 cm) at S$120 and the NVIDIA QSFP112 400G DAC (40 cm) at S$155. NVIDIA's Sync documentation specifies a QSFP112, 400 GbE, Ethernet-mode DAC and names the Amphenol NJAAKK-N911 and Luxshare LMTQF022-SD-R as supported; the link itself negotiates at 200 Gbit/s because that is the ConnectX-7's port speed on GB10. One cable carries full bandwidth. NVIDIA is explicit that a second cable does not improve performance and only means four interfaces to address. Mixing a 1 TB and a 4 TB unit is fine, but keep both on the same DGX OS release.
Placement and the physical link
A 40 cm cable dictates placement: the two units sit side by side with the rear panels facing the same way, which is also the right thing for cooling. Each GX10 draws up to 240 W through its own USB-C adapter and exhausts through the chassis, so leave both with clear airflow rather than stacking them. Connect the cable to the same physical QSFP port on each unit; NVIDIA's playbook says either port works but mismatching them causes NCCL test problems. Since January 2026 DGX OS supports hot-plugging the ConnectX-7, so you can cable running units. Keep both on the office switch over 10 GbE as usual; the QSFP link is a private back-to-back network for model traffic, not a replacement for your LAN.
Step 1: update both units
Every unit in the cluster must run the April 2026 DGX Spark software release or later; the Cluster Assistant checks this and refuses older systems. Update through the DGX Dashboard on each unit; ASUS listed DGX OS 7.5.0-2 for the GX10 on 29 July 2026, so anything current qualifies. Also make sure the same username exists on both units; NVIDIA's manual playbook creates a shared nvidia user with sudo useradd -m nvidia, sudo usermod -aG sudo nvidia and sudo passwd nvidia when the names differ.
Step 2: NVIDIA Sync Cluster Assistant (about 10 minutes)
NVIDIA Sync is the companion app for Windows, macOS and Ubuntu that manages SSH connections, tunnels and tools for a DGX Spark; download it from the Connect to your Spark page on build.nvidia.com and add both GX10s as devices. Then follow NVIDIA's documented sequence: open Settings > Cluster Assistant and choose Add New Cluster; name it; select the two devices; let the assistant check SSH access, GB10 hardware, minimum OS version and sudo; review usernames, UIDs and GIDs; run the network check, which blocks progress if a cable is missing; confirm the network plan; run the link speed test, which passes at a lower bound of 184 Gbit/s per link; wait for inter-device SSH setup (five-minute limit per machine); and copy the network information from the success screen. Each node can then reach the other with ssh <name>, because Sync writes aliases into your SSH configuration.
Manual alternative: netplan and SSH keys (about an hour)
If you prefer the terminal, NVIDIA's Connect Two Sparks playbook does the same job. Run ibdev2netdev on each unit; the interfaces for the cabled port show "(Up)", with two logical names for one physical port such as enp1s0f1np1 and enP2p1s0f1np1. Assign addresses persistently in /etc/netplan/40-cx7.yaml (node 1 takes 192.168.100.10/24 and 192.168.101.10/24, node 2 takes .11 on both; sudo chmod 600 the file, then sudo netplan apply), or temporarily with sudo ip addr add, remembering that temporary addresses vanish on reboot. Then run NVIDIA's discover-sparks.sh script from one node to distribute SSH keys, or use ssh-copy-id by hand. Rollback is deleting the netplan file and applying again.
Verifying the link
Three checks tell you the cluster is real. ibdev2netdev must show the port Up on both units; if not, NVIDIA's advice is to reseat the cable, reboot and retry. ip addr show enp1s0f1np1 must show the assigned address on each node. And ssh <IP of the other node> hostname must return the other machine's name without a password prompt. Sync users get a fourth check free: the link speed test, which should report close to 200 Gbit/s. NVIDIA also publishes a cluster network inspection guide and an NCCL bandwidth-test playbook for deeper checks.
Running one model across both units
NVIDIA's "Run on Two Sparks" vLLM playbook is the reference. Prerequisites on each unit: Docker with your user in the docker group, and the Hugging Face CLI installed and authenticated. On the head node clone eugr/spark-vllm-docker, download the model to every node with its hf-download.sh helper, then launch from the head node with launch-cluster.sh, which starts the vLLM container on both units and passes vllm serve the flag that matters, --tensor-parallel-size 2, so the model's layers are split across the two GPUs over the ConnectX-7 link. NVIDIA's worked example serves nvidia/Qwen3.8-27B-NVFP4 with FP8 KV cache and speculative decoding, and there is a one-line recipe variant, ./run-recipe.sh recipes/qwen3.8-27b-nvfp4-dflash2.yaml --setup. Test with a curl to http://localhost:8000/v1/chat/completions on the head node. Stop vLLM on both units before changing the cluster. For the 405B-class models that justify the second unit, the same tensor-parallel approach applies with a larger checkpoint; expect interactive rather than datacenter speeds, because traffic over the 200 GbE link adds latency a single 128 GB pool does not have. Our Ascent GX10 guide covers single-unit model tiers, and if the goal is a whole team hitting a 400B model at once, the ET900N G3 with 748 GB in one coherent pool is the better tool.
Three and four units
Three GX10s connect either as a direct ring (three cables, each unit to the other two) or through a switch; four units require a switch with a 200 Gbit/s Ethernet link for each device, and NVIDIA states that direct cabling is not supported at four. Do not mix direct and switched links in one cluster, and use one cable per link. The Cluster Assistant handles all of these topologies up to four devices; anything larger is a manual build. For the switch, confirm in its admin console that each port has negotiated 200 Gbit/s before running the Sync speed test.
Buying GX10 clusters in Singapore
SourceIT is an authorised ASUS reseller stocking all three Ascent GX10 configurations, both QSFP cables, and commissioning as a service: we deliver the pair updated, cabled, clustered through NVIDIA Sync and serving the model you choose, with a verified link test in the handover notes. GB10 units are supply-constrained, so order both units together to get matching firmware. Official ASUS Singapore warranty applies to every unit; education and NVIDIA Inception pricing may be available. Use Request for Quotation on the product page, call (65) 6978 3502, or email Sales@sourceit.com.sg; we are GeBIZ and Vendors@Gov registered. For where a cluster sits in ASUS's range see the ASUS Agent Computer guide, and for the other GB10 boxes our GB10 buyer's guide.
Frequently asked questions
Which cable do I need to link two GX10s? One QSFP112 400 GbE direct-attach cable in Ethernet mode. SourceIT sells the ASUS 40 cm cable at S$120 and the NVIDIA QSFP112 DAC at S$155; the link runs at 200 Gbit/s on GB10.
Do I need two cables for full bandwidth? No. NVIDIA states one cable achieves full bandwidth; a second adds nothing and complicates addressing.
Can I cluster a GX10 with an NVIDIA DGX Spark or HP ZGX Nano? The Cluster Assistant checks only for GB10 hardware and the minimum OS version, so mixed OEM units on the same DGX OS release can be clustered. Keep the firmware and OS matched.
Does clustering double the speed? No. It doubles memory. Tensor-parallel inference over the 200 GbE link is slower per token than a model that fits in one unit, so cluster for capacity, not throughput.
Do I need a switch for two units? No. Two units connect directly. A switch becomes mandatory at four units.
What DGX OS version is required? The April 2026 DGX Spark release or later on every unit. Update through the DGX Dashboard first.
Will SourceIT cluster them for me? Yes. Commissioning includes cabling, Cluster Assistant setup, the link speed test and serving your chosen model across both units.
