@MiaAI_lab: https://x.com/MiaAI_lab/status/2084629608062673379
Summary
A practical guide to selecting MikroTik switches (CRS504, CRS804, CRS812) for building DGX Spark clusters, including pricing and real user examples of 8x DGX Spark setups.
View Cached Full Text
Cached at: 08/04/26, 04:12 PM
The Ultimate DGX Spark Switch Guide
If you want to connect more than 3 DGX Sparks in a cluster, you’ll need a switch. Here’s a short, straight-to-the-point guide to choosing the right switch for your DGX Sparks cluster ✨
MikroTik CRS504
Planning only up to 4 DGX Sparks? Get the CRS504. Yes, it’s only 100GbE, but real-user reports show it delivers about 98% of the performance of much bigger and more expensive switches.
Retail: ~$759
MikroTik CRS804
Planning more than 4 Sparks or want to future-proof? Get the CRS804.
The CRS804 doesn’t natively have 8x 200G ports. Instead, it features four massive 400G QSFP56-DD ports. To connect 8 DGX Sparks at 200G, you must use breakout cables to split each 400G port into two 200G connections (4 ports × 2 = 8 nodes)
MikroTik CRS812
If you can’t find a CRS804 in stock, go with the CRS812.
To hook up 8 nodes to this switch, you would have to break down one of your 400G QSFP ports into four 100G links. This means you would end up with four nodes at 200G and four nodes at 100G**.**
Both CRS804 and CRS812 retail for ~$1,200
Real users who use a switch to connect 8x DGX Sparks
Here’s some actual users who run 8x DGX Spark clusters:
llmwish.open@ciprianveg·Jul 26GLM-5.2 can see now - 8× GB10 DGX Spark. Full instructions: https://github.com/ciprianveg/gb10-glm-5.2/tree/master/v18-vision… Prebuilt image: http://ghcr.io/ciprianveg/gb10-glm-5.2:v18-vision…
Credits: @MiaAI_lab (CosmicRaisins), https://huggingface.co/baseten https://huggingface.co/aidendle94 — the vision model ,overlay, adaptive MTP docker pull58826.2K
Takuya KAWATSU@takuz0_·Aug 1Testing 8x DGX Spark nodes in an open world setup :)373136447K
Gumbii.Digital@GumbiiDigital·Jul 25Just dropped the full thermal management write-up for the open-rack 8x DGX Spark cluster.
Left the rack open on purpose while I was still iterating. That created a heat pocket from the PSU bricks and cable loops sitting on the shelves.
Designed and 3D-printed custom supportsShow more211512810K
Similar Articles
Setting up of a 16xGB10 (DGX Spark) cluster
A user is setting up a 16-node DGX Spark cluster to run frontier open models locally, discussing networking tradeoffs between 200 and 100 Gbit/s.
@MiaAI_lab: What’s the best model you can run on your @NVIDIAAI DGX Spark? 1× DGX Spark * Qwen 3.6 35b NVFP4 - 256k ctx, 110 tok/s…
A tweet detailing the best AI models to run on Nvidia's DGX Spark, including Qwen 3.6 and DeepSeek v4 Flash variants, with token speeds and context lengths for single and multi-unit setups.
Strix Halo or DGX Spark for a home LLM server?
A user seeks recommendations on choosing between AMD Strix Halo and Nvidia DGX Spark hardware for setting up a local network-accessible LLM server.
DGX Spark agentic usage numbers
A user shares benchmark results and configuration for running Qwen3.6 models on NVIDIA DGX Spark using vLLM, focusing on agentic workloads with concurrent requests and tool calling.
@aijoey: for all my new dgx spark owners. https://github.com/joeynyc/spark-doctor…
Spark Doctor is an open-source diagnostic CLI for NVIDIA DGX Spark that collects system, GPU, memory, Docker, and recipe data, applies specific rules, and outputs the likely cause and next steps for common issues.