DGX Spark cluster: two Sparks, setup, limits and tok/s
Two DGX Sparks joined by one QSFP cable give you 256 GB of memory and, with tensor parallelism, faster decode as well. I measured 1.75× for a single request on Qwen3-32B-AWQ with vLLM tensor parallelism on a GPUwerk pair (12 November 2026). Owners report up to about 1.8× on dense models with vLLM or SGLang, and about 44 tok/s for DeepSeek-V4-Flash at FP8 on two Sparks. NVIDIA says a pair handles models up to 400B parameters. You can rent a cabled pair on GPUwerk at $1.79/hour, billed by the minute.
| Question | Answer | Source |
|---|---|---|
| Memory | 256 GB (2 × 128 GB) | NVIDIA DGX Spark product page |
| Largest model | Up to 400B parameters on 2 Sparks, up to 200B on 1 | NVIDIA DGX Spark product page |
| Decode speed-up, measured | 1.75× for one request, 1.63× at 16 concurrent, Qwen3-32B-AWQ, vLLM 0.33.0, TP=2 (GPUwerk-measured) | GPUwerk benchmarks |
| Decode speed-up, reported | Up to about 1.8× on dense models, vLLM or SGLang, tensor parallel (reported) | NVIDIA forum thread 361862 |
| Link | One QSFP cable, up to 200 Gb/s per port, Ethernet only | NVIDIA DGX Spark clustering docs |
| Setup time | About 10 minutes with NVIDIA Sync, about 1 hour by hand | NVIDIA playbook |
Does a second Spark make decode faster?
Yes, when the model is split with tensor parallelism. Decode on a Spark is limited by memory bandwidth: each new token means reading the model's active weights from memory. Tensor parallelism (TP=2) splits every layer's weights between the two nodes, so each node reads only half of them per token. Both nodes read at the same time, so per-user decode can rise. The cost is a small exchange of activations over the cable at every layer, which is why the gain stays below 2×.
Pipeline parallelism and llama.cpp's RPC mode are different. They give each node whole layers in turn, so the nodes mostly wait for each other. That fits a bigger model, but it does not make one user's decode faster.
Measured on a GPUwerk pair
On 12 November 2026 I rented a two-Spark cluster through the GPUwerk console (a partner-hosted pair in the United States, one ConnectX-7 cable) and served Qwen3-32B-AWQ, a dense 32B model in 4-bit, with vLLM 0.33.0 on one node and then across both with TP=2. The workload was vllm bench serve with random prompts, 1,024 tokens in and 256 out.
| Concurrency | One Spark, generated tok/s | Two Sparks, generated tok/s | Speed-up |
|---|---|---|---|
| 1 | 11.47 | 20.10 | 1.75× |
| 4 | 42.12 | 70.65 | 1.68× |
| 16 | 101.32 | 164.84 | 1.63× |
GPUwerk-measured, one run per row, one model, one pair, one day; not a guarantee for your model or workload. The KV cache grew from 338,112 tokens on one node to 731,392 across two. NCCL all_reduce over the cable reached 22.1 GB/s (177 Gbit/s) at 1 GiB. Full table, commands and caveats: two Sparks measured on the benchmarks page, with the raw vLLM output.
Reported two-Spark results
These are owners' reports on NVIDIA's developer forum, not GPUwerk measurements. Results move with the engine version, the model build and the context length.
| Model and quantisation | Engine and split | Reported result | Source |
|---|---|---|---|
| Dense models (general) | vLLM or SGLang, tensor parallel, 2 Sparks | Up to about 1.8× a single Spark | Forum 361862, reply by eugr |
| DeepSeek-V4-Flash, official FP8 (about 149 GB of weights), FP8 KV cache | vLLM (jasl fork, eugr's spark-vllm-docker), TP=2, 2 Sparks, 200K context | About 44 tok/s decode, warm, single stream; about 45 tok/s total at 2 concurrent streams | Forum 370309, 16 May 2026 |
| DeepSeek-V4-Flash, same setup, other owners | vLLM, TP=2, 2 Sparks | 32.7 to 37.3 tok/s decode in their own benchmark runs | Forum 370309, replies of 17 May 2026 |
| Qwen3.5-397B-A17B, int4 AutoRound | vLLM, TP=2, 2 Sparks | 26 to 30 tok/s for one user; 42 to 46 tok/s total for two | Forum 361967, 28 February 2026 |
| Qwen3-235B-A22B, NVFP4 (nvidia/Qwen3-235B-A22B-FP4) | TensorRT-LLM, TP=2, 2 Sparks | 21.65 tok/s (trtllm-bench, release 1.2.0rc5) | Forum 355053, December 2025 |
These rows are third-party reports, not GPUwerk measurements; my own run is in the table above. Read 24 October 2026. Single-node figures for the same models are on the benchmarks page.
Which models need two Sparks?
A model needs two Sparks when its weights plus the KV cache for your context do not fit in one Spark's 128 GB. Weight sizes below come from the linked sources, or from parameters × bytes per parameter where marked.
| Model | Format | Weights | One Spark? | Two Sparks |
|---|---|---|---|---|
| DeepSeek-V4-Flash | Official FP8 | About 149 GB (forum 370309) | No, over 128 GB | Runs with TP=2 at 200K context, about 44 tok/s reported |
| Qwen3.5-397B-A17B | int4 AutoRound | About 200 GB (397B × 0.5 bytes, estimate) | No | Runs with TP=2, 26 to 30 tok/s reported |
| Qwen3-235B-A22B | FP8 | About 235 GB (235B × 1 byte, estimate) | No | Tight: little room left for KV cache |
| Qwen3-235B-A22B | NVFP4 or 4-bit GGUF | 134 GB for UD-Q4_K_XL (forum 355126) | Only just, with a short context | Comfortable; 21.65 tok/s reported with TensorRT-LLM TP=2 |
| Any model of 200B to 400B parameters | FP4 | About 100 to 200 GB | Up to about 200B | Up to about 400B, per NVIDIA |
Rows marked "estimate" are my arithmetic on weight bytes only. Real memory use is higher once the KV cache and runtime are loaded.
What cable and parts do I need?
One short QSFP direct-attach cable between the two Sparks. No switch is needed for two or three units. NVIDIA's clustering documentation lists the details below (read 24 October 2026).
| Part | What NVIDIA specifies |
|---|---|
| Cable | QSFP to QSFP112 direct-attach copper, rated for at least 200 Gb/s. A faster cable gives no gain because the port is capped at 200 Gb/s. |
| Approved cables | Amphenol NJAAKK-N911 (400 mm, 32 AWG; NJAAKK0006 is the 0.5 m version) and Luxshare LMTQF022-SD-R (QSFP112 400G DAC, 400 mm) |
| Ports | Two QSFP ports per Spark on the ConnectX-7 NIC, up to 200 Gb/s each, Ethernet only. The left port is the one next to the RJ45 Ethernet port. |
| Interfaces per port | Each QSFP port shows up as two Linux interfaces, because the NIC connects through two PCIe Gen 5 x4 links |
| Left port (0) | enp1s0f0np0 and enP2p1s0f0np0 |
| Right port (1) | enp1s0f1np1 and enP2p1s0f1np1 |
| Interface lookup | Run ibdev2netdev. It maps each RoCE device to its interface and shows (Up) on a cabled port. Names are the same on every Spark. |
| Switch (4 Sparks) | A 200 GbE switch with QSFP56-DD ports and RoCE v2 support, per NVIDIA's product page |
How do I set up two Sparks?
Follow NVIDIA's playbook Connect Multiple DGX Sparks for Distributed Workloads. It has two paths:
- NVIDIA Sync. The Cluster Assistant sets up the ConnectX-7 network and tests it. NVIDIA puts this at about 10 minutes.
- Manual. NVIDIA's
spark_cluster_setupscript configures the network and SSH between the nodes. NVIDIA puts this at about 1 hour and expects you to know Netplan and IP addressing.
Then serve a model across the pair with NVIDIA's Serve LLMs with vLLM Across DGX Sparks playbook, and check the link with the NCCL playbook. Before you start, confirm both cabled interfaces show (Up) in ibdev2netdev. Commands and flags change between DGX OS releases, so use NVIDIA's current pages rather than a copy from a forum post.
On my own test pair, at the time of the test, pip-installed vLLM also needed the Python headers (python3.12-dev) and the CUDA 13 toolkit on PATH, and the two-node run needed VLLM_HOST_IP, NCCL_SOCKET_IFNAME and GLOO_SOCKET_IFNAME set to the fabric interface. The exact commands are on the benchmarks page.
Limits of a two-Spark cluster
- Below 2×. Each layer exchanges activations over a 200 Gb/s link, far slower than each Spark's own memory, so tensor parallelism gains less than double.
- Engine matters. The speed-up above is for vLLM or SGLang tensor parallelism. llama.cpp RPC fits larger models but adds little or no speed (forum 361862).
- Port count. NVIDIA Sync supports up to three Sparks cabled directly; four need a switch.
Should you pair two Sparks?
Pair them if your model does not fit in 128 GB, or if you want more single-user speed from a model that a TP=2 engine supports. Stay on one Spark if your model fits and your engine has no tensor-parallel mode across nodes. The fine-tuning guide covers fine-tunes that need a pair, and the DGX Spark vs Mac Studio page compares single-node speed. For three to eight Sparks behind a switch, see DGX Spark clusters of up to eight nodes.
Test before you buy a second box. Rent a DGX Spark on its own at $0.79/hour, or a cabled pair at $1.79/hour, both billed by the minute.
FAQ
Does a second Spark make it faster?
Yes, with tensor parallelism in vLLM or SGLang. Each node reads half the weights per token. On a rented GPUwerk pair I measured Qwen3-32B-AWQ at 20.10 tok/s for one request across two Sparks against 11.47 on one, 1.75× (vLLM 0.33.0, 12 November 2026), and owners report up to about 1.8× on dense models. DeepSeek-V4-Flash at FP8 is reported at about 44 tok/s on two Sparks with TP=2. llama.cpp RPC mostly adds memory, not speed.
What cable do I need?
One QSFP to QSFP112 direct-attach cable rated for at least 200 Gb/s. NVIDIA lists the Amphenol NJAAKK-N911 and the Luxshare LMTQF022-SD-R, both 400 mm. No switch is needed for two Sparks.
How much memory does a two-Spark cluster have?
256 GB, two times 128 GB. NVIDIA says a pair runs models up to 400B parameters, against up to 200B on one Spark.
How do I find the interface names for the cable?
Run ibdev2netdev. The left QSFP port is enp1s0f0np0 and enP2p1s0f0np0, the right port is enp1s0f1np1 and enP2p1s0f1np1, and the names are the same on every Spark.
Can I rent a two-Spark cluster?
Yes. GPUwerk rents a cabled pair of DGX Sparks at $1.79/hour, billed by the minute.
Sources, read 24 October 2026: NVIDIA DGX Spark clustering docs; NVIDIA DGX Spark product page; NVIDIA playbooks connect-multiple-sparks and vllm-cluster; NVIDIA developer forum threads 361862, 370309, 361967, 355053 and 355126. GPUwerk rates: pricing.
A first engagement can help scope whether your workload needs a second node before you cable one up.