How many DGX Spark you can connect, and what one Spark does beyond language models

Why link two Sparks: 256 GB for larger models

There is no DGX Spark with 256 GB of memory. NVIDIA lists the system with 128 GB or, through OEM partners, 64 GB. Two 128 GB Sparks linked by one QSFP cable give 256 GB, and NVIDIA rates them for models of up to 400 billion parameters. Each Spark keeps its own 128 GB, so the model is split across the two over the link, and with the memory settings in NVIDIA’s playbooks roughly 204 to 230 GB remain for weights and KV cache. Which models need 256 GB is worked through in our DGX Spark 256 GB article.

As our first DGX Spark article shows, 273 GB/s of memory bandwidth limits single-user generation on a dense 70B model to a few tokens per second. A second Spark can raise that limit only through tensor parallelism, and the link between the two Sparks takes back some or all of the gain. It also doubles the memory: two 128 GB Sparks hold 256 GB, so models that do not fit on one Spark, from a 70B in FP16 to a 235B mixture-of-experts in NVFP4, can be loaded. NVIDIA promotes linking two Sparks mainly for this capacity. Two of the 64 GB systems that OEM partners have sold since 23 October 2026 pool 128 GB, which NVIDIA rates for models of up to 200 billion parameters.

How the QSFP link works

Each Spark has two QSFP ports on its ConnectX-7 controller. The port is rated at 200 Gbit/s; the controller connects to the GB10 chip through two independent PCIe Gen5 x4 links, and each QSFP port appears in Linux as two Ethernet interfaces, each with its own RoCE device. NVIDIA’s own performance guide measures 92.57 and 97.28 Gbit/s across the two logical links, 189.85 Gbit/s combined. Plan the connection as two 100G links rather than one 200G link; the Cluster Assistant checks for a lower bound of 184 Gbit/s and flags anything below it.

One approved QSFP112 direct-attach cable between the two Sparks is all the wiring needed; NVIDIA states that full bandwidth is reached with a single cable. Two or three Sparks can be cabled directly; four need a switch with one 200 Gbit/s Ethernet link for each Spark, as NVIDIA’s Connect-multiple playbook puts it. The NVIDIA Sync Cluster Assistant supports two to four Sparks and runs the link checks and the speed test, but it does not set up the workload. NVIDIA’s blog of 21 October 2026 says DGX Spark “supports clusters of up to eight units in manual configuration” and that NVIDIA Sync is to support eight Sparks later in October 2026; the product page and the clustering guide, both updated on 21 October 2026, still name four.

Our engineering partner Vixen.UNO can also set up the workload across the Sparks, under the same contract as the hardware. Write to us with the model you want to split across them; the first consultation on setup and integration is free of charge.

What NVIDIA supports, and what it claims

NVIDIA’s playbooks specify the software, and NVIDIA says the Model Launcher in NVIDIA Sync, due later in October 2026, downloads a model and container suited to one Spark or a cluster. For vLLM, NVIDIA Sync sets up the cluster and a recipe run from the head Spark serves the model across both Sparks with a tensor-parallel size of two. TensorRT-LLM uses OpenMPI with the same tensor-parallel size and the Qwen3-235B-A22B-FP4 recipe. For training, NVIDIA’s January 2026 update describes distributed fine-tuning of models up to 70B parameters across two Sparks with FSDP and LoRA.

NVIDIA gives two figures for two Sparks: its product page, updated on 21 October 2026, says up to 400 billion parameters on two 128 GB systems and up to 700 billion on four, while its hardware overview of the same date still says 405 billion for a dual-Spark configuration. These parameter counts describe memory capacity rather than speed.

Measured two-node numbers

The two-node link is slow enough that pipeline parallelism, which sends activations across the link once per pipeline stage, outperforms tensor parallelism, which synchronises at every layer: 555 against 252 tokens per second on the same model at batch 128. Single-stream generation on a 235B or 320B model reaches 12 to 15 tokens per second without speculative decoding, which suits one engineer. DevelopersIO measured 76.2 tokens per second in total across eight parallel requests on the 320B model, with multi-token prediction. A second Spark adds memory capacity first, and speed per request at most in part.

Beyond language models

Image generation and fine-tuning depend on compute more than on memory bandwidth, so the GB10’s fifth-generation Tensor Cores matter more there than its 273 GB/s.

Image generation is fast. A 1K FLUX image every 2.6 seconds is a workable pace for a design team, and with 128 GB the FP16 weights of the large diffusion models load without offloading. Video generation is slower. The desktop RTX 5090, with a fraction of the memory, finishes the same Hunyuan clip in a third of the time, because video generation is compute-heavy and the GB10 has roughly a quarter to a third of the desktop card’s Tensor throughput, as well as laptop-class memory bandwidth.

Fine-tuning falls between the two. QLoRA on a 70B model at 760 tokens a second is slow by data-centre standards but practical for an overnight run on a few thousand documents, which is the scale of most corporate fine-tuning. Robotics is supported in the sense that Isaac Sim and Isaac Lab build from source for aarch64 and NVIDIA publishes a playbook; it is not a click-to-run environment.

Under a mainly GPU-driven load Tom’s Hardware measured about 160 W at the wall; NVIDIA declares 233.2 W maximum and 38.0 W idle at 230 V. The January 2026 software update added hot-plug support for the ConnectX-7 adapter, which NVIDIA says saves up to 18 W while the network ports are unused. Tom’s Hardware, measuring at the USB-C input, saw idle power fall from about 37 W to between 22 and 25 W after that update.

Gaming and Windows on DGX Spark

DGX OS is Ubuntu on Arm. There is no Windows, and we found no NVIDIA statement that games are supported on the product. Games have been run through x86 emulation. Tom’s Hardware reported a user’s test of Cyberpunk 2077 at around 50 frames per second at 1080p medium through Box64 and Steam, and a post on NVIDIA’s developer forum shows 75 frames per second at 1080p low with Canonical’s Steam snap, FEX and Proton. Guru3D reported a setup guide shared by NVIDIA staff on the forums that passes 175 frames per second with DLSS 4 multi-frame generation. None of these setups is supported by NVIDIA, and a driver update can break them.

For Windows, NVIDIA has a different product. RTX Spark was shown at Computex 2026; its largest version has the same 6,144 CUDA cores, a 20-core Grace CPU and up to 128 GB, and NVIDIA lists Windows 11 as its operating system. NVIDIA’s press release of 16 October 2026 says “RTX Spark laptops are available globally today”, and its developer blog says the small form factor desktops are “coming in November”. Our comparison of DGX Spark and RTX Spark sets out the differences; DGX Spark is built as a development system for models.

What DGX Spark is for, and when a GPU server is needed

NVIDIA still positions DGX Spark for local inference on models up to 200 billion parameters and fine-tuning of models up to 70 billion, for development and validation; its product page gives the 200 billion for the 128 GB system and 100 billion for the 64 GB configuration. Two Sparks extend the inference limit and make the fine-tuning limit comfortable. Neither replaces a GPU server when the model goes into production in front of users, because of the bandwidth figures above. When the model moves to production, our engineering partner Vixen.UNO selects the GPU servers for it, from a single server to a cluster, as our Private AI/ML page describes, and we supply them as AI servers built to order.

What we supply

We supply the NVIDIA DGX Spark Founders Edition (128 GB, 4 TB) across the EU on one contract and invoice, with manufacturer warranty, for example two Sparks to link into 256 GB or four for a team; quantities and lead time are confirmed in writing with the quote.

FAQ