Build Daily
Tinley Park · October 21, 2026DGX Spark Dell GB10 2 Node Cluster
The video documents building an 8-node DGX Spark (Dell GB10) cluster with 1 TB of combined VRAM, going beyond NVIDIA's official two-node guidance to enable local large-language-model inference. It covers the hardware, 400 Gbps networking, and software stack used to make the cluster work.
The outline
Every point the video makes, in the order it makes them.
- 01
8-Node Cluster Build
The creator wires eight DGX Spark units together to form a single 1 TB VRAM cluster, a configuration NVIDIA never officially demonstrated.
- 02
Beyond the 2-Node Limit
NVIDIA only showed how to link two DGX Sparks, so the video fills the gap by scaling the setup to eight nodes.
- 03
400 Gbps Networking
High-speed 400 Gbps switches (2-port and 4-port variants) are used to interconnect the nodes for fast data transfer across the cluster.
- 04
LLM Inference on the Cluster
The assembled cluster is intended to run large language models locally, such as 70 B-parameter models, leveraging the pooled VRAM.
- 05
vLLM / Docker Software Stack
A community GitHub project (eugr/spark-vllm-docker) provides the Docker-based vLLM setup needed to serve models across the multi-node Spark cluster.
- 06
DGX Spark Giveaway
NVIDIA sponsors a DGX Spark giveaway for viewers who attended GTC 2026, with entries open March 15–26, 2026.
Eight nodes, 1 TB combined VRAM, 400 Gbps networking — a local inference stack that pushes past NVIDIA's two-node guidance. For a solo builder, removing the per-token cloud bill changes what is economically shippable. The path there is physical and fiddly: USB-C KVM adapters, a knife for cable strain relief, update cycles that restart the box before the dashboard agrees it is done.