DGX Spark 256 GB: there is no such model, and how two Sparks give you 256 GB

Is there a DGX Spark with 256 GB

There is no DGX Spark with 256 GB of memory. NVIDIA’s product page lists 64 GB and 128 GB configurations, and its footnote states that the “64 GB memory configuration is available exclusively through participating OEM partners”; as of 16 October 2026 the page marks it “Coming Soon”, and NVIDIA’s blog dates the OEM systems to 23 October 2026. The Founders Edition has 128 GB. The 256 GB figure is the memory of two 128 GB Sparks linked by one QSFP cable, a setup the same page shows as “2x DGX Spark” with 256 GB and “Up to 400B parameters”.

The 128 GB is LPDDR5x on a 256-bit interface at 273 GB/s, and the CPU and the GPU share it, so on a DGX Spark the RAM and the VRAM are the same pool, and 256 GB of RAM also means two linked Sparks. Our DGX Spark specifications list the remaining figures. We found no single GB10 system, from NVIDIA or from a partner manufacturer, with more than 128 GB, as of 16 October 2026.

Some resellers list a “bundle” of two DGX Spark systems with a connecting cable. NVIDIA’s product page shows the 2x and 4x configurations as ways to combine standard 128 GB systems, not as separate products.

One, two or four Sparks: memory, link and support

Each Spark keeps its own 128 GB. Linked Sparks do not form one memory pool that any program can address; a model is split across them by the serving software, layer by layer or tensor by tensor, and the parts exchange data over the network link.

NVIDIA DGX Spark product page and user guide (hardware overview and clustering chapter, updated 10 September 2026), read on 16 October 2026; weights and cache are our arithmetic, 0.8 and 0.9 of the total memory, the settings in NVIDIA’s vLLM playbook.

The 0.8 and 0.9 are the --gpu-memory-utilization values in the two-Spark examples of NVIDIA’s vLLM playbook (GitHub version, last updated 12 June 2026). Applied to 128 GB, the same fractions give the 102 to 115 GB that our article on what fits on one DGX Spark uses; NVIDIA’s single-Spark launch script sets 0.7, about 90 GB. The figures are estimates for weights and KV cache together; the operating system keeps the rest.

NVIDIA’s parameter counts are capacity figures, and by our arithmetic they hold only for 4-bit weights. At 4.5 bits per weight, 400 billion parameters take about 225 GB of the 256 GB in two Sparks, and 700 billion take about 394 GB of the 512 GB in four.

How many DGX Spark can you connect

NVIDIA’s clustering documentation states that the platform “supports up to three DGX Spark systems connected directly through cables, and up to four systems when using a switch”. The product page gives the same limit of four. NVIDIA’s switch playbook asks for a “QSFP switch with at least 4 QSFP56-DD ports (at least 200Gbps each)” and one cable per Spark, and adds that “the same set up can be scaled up to more DGX Spark devices connected via the same switch”. The product page and the clustering documentation still name four as the maximum, so plan with four.

Two Sparks need one cable. NVIDIA’s two-Spark playbook states that “full bandwidth can be achieved with just one QSFP cable” and asks for the same physical port on both systems. Each port appears in Linux as two logical interfaces. The clustering chapter names the approved direct-attach cables, the Amphenol NJAAKK-N911 and NJAAKK0006 and the Luxshare LMTQF022-SD-R, and notes that a faster cable gains nothing “because the port itself is capped at 200 Gb/s”. As our guide to buying a DGX Spark in the EU notes, the cable is not in the box, so plan one approved cable per link.

What 256 GB on two Sparks costs in speed and setup

Inside one Spark, the GPU reads its memory at 273 GB/s. Between two Sparks, the port is rated at 200 Gb/s, about 25 GB/s, roughly one eleventh of that. A model split across two Sparks therefore runs at a pace set by both figures. With tensor parallelism, each Spark reads half of the weights for every token, which raises the bandwidth ceiling, and the two exchange data over the link at every layer.

NVIDIA’s own measurement shows the effect. In its technical blog of 24 October 2025, Qwen3 235B in NVFP4 on two Sparks with TensorRT-LLM generated 11.73 tokens per second at batch size 1, with 23,477 tokens per second of prompt processing (2,048-token prompt, 128 generated tokens). The model has 22 billion active parameters, about 12.4 GB per token at 4.5 bits. Split across two Sparks, that is about 6.2 GB read on each, a bandwidth ceiling of roughly 44 tokens per second by our arithmetic. The measured rate is about a quarter of it. For comparison, gpt-oss-120b, which fits on one Spark, ran at 55.37 tokens per second with llama.cpp in the same NVIDIA table.

The setup is documented step by step and takes some effort. NVIDIA’s two-Spark playbook estimates “1 hour including validation” for the network link alone. The Cluster Assistant in NVIDIA Sync “validates the devices, applies ConnectX-7 network settings, checks link performance, and configures SSH between nodes”. The serving engine is set up after that, and NVIDIA’s vLLM recipe runs Llama 3.3 70B across two Sparks with a tensor-parallel size of 2 and the Ray backend. Our two-node article covers the wiring, the commands and further measured results.

Which models need 256 GB

A model needs a second Spark when its weights plus the KV cache for your context and users exceed about 102 to 115 GB. Weight size is parameters times bytes per parameter: 2 for BF16, 1 for FP8 and about 0.56 for NVFP4 with its block scales.

Parameters and sizes from the model cards and file lists on Hugging Face (NVIDIA’s Qwen3-235B-A22B-FP4 checkpoint, MiniMax-M2.5, the 405B AWQ-INT4 checkpoint) and from OpenAI’s gpt-oss-120b checkpoint, read on 16 October 2026; Spark budgets as in the first table; GPU servers by our sizing rule, 0.9 of the memory the driver reports minus 3 GiB, about 89 GB per RTX PRO 6000 and 132 GB per H200 NVL.

The 405B row shows the limit of two Sparks. NVIDIA’s vLLM playbook loads the 405B AWQ checkpoint on two Sparks with --max-model-len 64, a single sequence and 0.9 of memory, and warns that the “405B model has insufficient memory headroom for production use”. The checkpoint needs about 203 GiB to load, according to its model card, which leaves about 12 GB of the 230 GB for cache, activations and CUDA graphs. On four Sparks the checkpoint loads with room for cache, but as a dense model it reads all of its weights for every token, about 54 GB per Spark, which by our arithmetic limits one request to about 5 tokens per second. For MiniMax-M2.5, 229 billion parameters in FP8, NVIDIA’s switch example uses a tensor-parallel size of 4 and notes: “If you are running with four or more sparks, you can comfortably run this model with tensor parallelism.”

Context adds to every row. Qwen3-235B-A22B has 94 layers, 4 KV heads and a head dimension of 128, so one 32,768-token conversation takes about 3.15 GB of KV cache in FP8. NVIDIA’s NVFP4 checkpoint of the model takes 134 GB, so the 71 to 96 GB left on two Sparks hold about 22 to 30 such conversations, and half as many with a 16-bit cache.

When one Spark is enough

One Spark is enough when the model and its cache fit in about 102 to 115 GB. That covers gpt-oss-120b, a 70B model in FP8 or NVFP4 and 4-bit models well below 200 billion parameters. A 4-bit model at NVIDIA’s figure of 200 billion fits only with little room for cache. NVIDIA’s TensorRT-LLM playbook lists one model in its support matrix as “two Sparks only”, Qwen3-235B-A22B, so the limit of one Spark lies below the 235B class.

For a team that wants more simultaneous users on a model that already fits, two Sparks can each serve their own copy behind a load balancer instead of being linked. Each copy runs at the speed of one Spark, and no request waits for traffic over the link. Our article on one DGX Spark for a team puts one Spark at about 16 simultaneous gpt-oss-120b requests with 4,096-token prompts at 10 tokens per second or more, by the llama.cpp maintainers’ results.

When an RTX PRO 6000 or H200 NVL server fits better

Linked Sparks hold large models at a pace suited to one developer or a batch job. When a 235B or 405B model has to answer a whole department, memory bandwidth decides the outcome. An RTX PRO 6000 Server Edition moves 1,597 GB/s and an H200 NVL 4.8 TB/s per card, against 273 GB/s per Spark. The H200 NVL also links 2 or 4 cards over an NVLink bridge at 900 GB/s per GPU, while the RTX PRO 6000 exchanges data over PCIe.

Use the Sparks to choose and test the model, then size the server from that model and the peak of requests in flight. The last column of the table above gives the smallest card count by our sizing rule, rounded up to 1, 2, 4 or 8 cards, the usual tensor-parallel sizes; more users need more cache and more cards.

We supply the DGX Spark and AI servers built to order with RTX PRO 6000 Server Edition or H200 NVL cards. Tell us the model, its precision and your context length, and we reply within one business day with a configuration and a quote.

Worked example: a team of 10 to 20 with two or four Sparks

A team of 10 to 20 developers and data scientists wants a chat and coding assistant for daily work and a 235B model to evaluate before a decision on a GPU server. Size it by the peak of requests at once from a pilot’s gateway logs, not by headcount.

With four Sparks, two are linked over one QSFP cable as a 256 GB system that runs Qwen3-235B-A22B in NVFP4 for the evaluation, at NVIDIA’s measured 11.73 tokens per second for one request. The other two serve gpt-oss-120b as separate copies behind a load balancer, about 32 requests at once with 4,096-token prompts, each at 10 tokens per second or more. Four Sparks draw at most about 0.9 kW from desk sockets, by NVIDIA’s declared 233.2 W maximum each.

With two Sparks, both serve gpt-oss-120b for daily work and are linked for the 235B evaluation outside working hours. Switching costs a restart of the serving engines, and during the evaluation the team has no assistant. If the evaluation shows that the team needs the 235B model every day, the next step is a server with two or more RTX PRO 6000 Server Edition cards, sized as above.

We supply two or four Sparks on one EU contract and invoice, with manufacturer warranty. Send us the team size, the peak of requests and the models, and we reply within one business day with a configuration and a quote.

What we supply

For 256 GB we supply two NVIDIA DGX Spark Founders Edition systems (128 GB, 4 TB each), and four where a team needs 512 GB or separate copies, across the EU on one contract and invoice with manufacturer warranty; quantities are confirmed in writing with the quote. When a model outgrows linked Sparks, we build AI servers to order with the RTX PRO 6000 Server Edition or the H200 NVL, and we check the rack, power and airflow before we quote. Setting up the serving engine across the Sparks or on the server is part of our Private AI/ML service, with engineering by our partner Vixen.UNO.

FAQ