Not quite, and sometimes not at all. What two linked NVIDIA GB10 systems reliably double is memory: two 128 GB units pool 256 GB, enough for models of about 400B parameters that one unit cannot load. Whether speed improves depends on how the model is split. With tensor parallelism in vLLM, published tests show one user's generation speed on a dense model rising by about 1.8 times on two units. With llama.cpp's simpler split, testers found no speed gain. And if many people are using a model that already fits on one unit, two separate units serve more requests than two clustered ones. Two ASUS Ascent GX10 1 TB units cost S$16,200 at SourceIT, plus a S$155 NVIDIA QSFP112 400G DAC cable (listed 28 Oct 2026).
What do published two-unit tests show?
| Test | One unit | Two units | Change | Source |
|---|---|---|---|---|
| Qwen3.6-27B NVFP4, vLLM tensor parallel, one request | 12.6 tok/s | 22.6 tok/s | About 1.8x | Openzeka, Jul 2026 |
| Qwen3.8-27B NVFP4, NVIDIA test, one request | 11.9 tok/s | Not published for 2 (4 units with switch: 36.3) | About 3x at 4 units | NVIDIA, Oct 2026 |
| DeepSeek-V4-Flash, llama.cpp split across units | Baseline | About 0.97x generation, 0.85x prompt | No gain | Classmethod, Aug 2026 |
| Qwen3-235B NVFP4, TensorRT-LLM (does not fit on one) | Does not fit | 11.73 tok/s generation, 23,477 tok/s prompt | New capability | NVIDIA, Oct 2025 |
| Qwen3.6-27B, 64 users, total throughput (4 units) | Four separate units: about 1,203 tok/s. One 4-unit cluster: about 508 tok/s | Separate wins | Openzeka, Jul 2026 | |
Third-party and NVIDIA results with different models and settings; read them as patterns rather than a direct comparison. We have not run these tests ourselves.
Why does the split method matter so much?
There are two common ways to run one model across two machines.
- Tensor parallelism splits every layer, so both GPUs work on every token at the same time. Each GPU reads half the weights, which roughly doubles the memory bandwidth working on a token, and bandwidth is what limits generation speed. The cost is that the two units must exchange data for every layer over the ConnectX-7 link. vLLM uses this approach, and Openzeka's 1.8x result shows it works well on a 27B model.
- Pipeline or layer splitting gives the first half of the layers to one unit and the second half to the other. Each token passes through one unit and then the other, so only one GPU works on it at a time. It adds capacity, not speed. llama.cpp's multi-machine mode works this way, which explains Classmethod's 0.97x.
NVIDIA's own clustering guidance, and our step-by-step guide to clustering two GX10s, use vLLM with tensor parallelism across two units.
Clustering is a way to run a bigger model, not a cheaper way to run a faster one.
Why not use two units for more users?
Because the link between them becomes the bottleneck. When a model fits on one unit, running a separate copy on each and splitting users between them avoids sending data across the network for every token. Openzeka's July 2026 test makes the point with four units: separate copies served about 1,203 tokens per second in total to 64 users, a single four-unit cluster about 508. For a team, two separate GB10s behind a simple load balancer is usually the better design. Our post Can One GB10 Support 20 Employees Using AI at the Same Time? covers sizing.
When does a second unit make sense?
- The model does not fit in 128 GB. Models between about 200B and 400B parameters at 4-bit, such as Qwen3-235B, need two units. NVIDIA rates two units for models up to 400B.
- You want faster single-user speed on a dense model and will use vLLM with tensor parallelism, accepting the setup work.
- You are preparing for data-centre deployment and need to test multi-node code on cheap hardware.
- You need redundancy, in which case run two separate copies, not a cluster.
What do you need to link two units?
| Item | Notes | SourceIT price |
|---|---|---|
| Two GB10 systems | Same DGX OS release on both; mixing brands is possible but matching is simplest | ASUS GX10 1 TB S$8,100 each; DGX Spark S$8,800 each |
| One QSFP112 DAC cable | NVIDIA says one cable gives full bandwidth; the port runs at 200 Gb/s | NVIDIA QSFP112 400G DAC S$155; HP ZGX QSFP112 DAC S$365 |
| Software | NVIDIA Sync Cluster Assistant, April 2026 system software or later | Included |
| Switch | Not needed for two units | None |
Prices listed 28 Oct 2026. NVIDIA lists approved cables by manufacturer part number; we confirm compatibility when quoting. Going beyond two units is covered in Four GB10 Supercomputers on One Desk, and the link itself in NVIDIA ConnectX-7 vs Ordinary Ethernet.
Buying in Singapore
SourceIT supplies GB10 systems from ASUS, NVIDIA, HP, Dell, Lenovo, GIGABYTE and MSI, the cables to link them, and commissioning: we can deliver a pair updated, cabled and serving your chosen model. Order both units together so they arrive on matching software. Browse the NVIDIA DGX Spark collection or the Ascent GX10 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
Does linking two DGX Sparks double performance? It doubles memory to 256 GB. Speed depends on the method: vLLM tensor parallelism raised one user's speed on a 27B model by about 1.8x in Openzeka's tests, while llama.cpp's split gave no gain in Classmethod's.
Is it better to cluster two GB10s or run them separately? For models that fit in one unit and many users, run them separately. Cluster when the model needs more than 128 GB or you want faster single-user speed with tensor parallelism.
What is the biggest model two GB10s can run? NVIDIA rates two units for models up to about 400B parameters. NVIDIA's own example ran Qwen3-235B at 4-bit at 11.73 tokens per second.
Do I need a switch for two units? No. Two units connect directly with one QSFP cable. A switch is required at four units.
What does a two-unit setup cost in Singapore? At SourceIT on 28 Oct 2026, two ASUS Ascent GX10 1 TB units are S$16,200 and an NVIDIA QSFP112 DAC cable is S$155.
