AI BriefingBuilding a 16-Node DGX Spark Cluster
Key point
A cluster of 16 DGX Sparks runs at line rate on the FS N8510 fabric.
Details
The build connecting 16 DGX Sparks to an FS N8510 fabric switch is complete, and the whole setup runs at line rate.
Each node shipped with the Nvidia version of Ubuntu pre-installed, and after rack mounting, the process was completed in order: creating identical accounts, waiting about 20 minutes per node for updates, passwordless SSH, jumbo frames, and IP configuration.
Each node uses 1x QSFP56, and the DGX Spark bonds two NICs into a single port to create a dual rail. Actual measurements show 100–111Gbps per rail, totaling around 200Gbps.
The goal is to secure a large amount of unified memory capacity to serve models.
- Served GLM-5.1-NVFP4 (434GB) on 8 nodes with TP=8.
- Next test targets are DeepSeek and Kimi.
- Long term, the goal is a prefill/decode separation architecture, with plans to add 2-4 M5 Ultra Mac Studios for decode.
- The rack also houses a QNAP 374TB U.2 NAS, 2 Dual 4090 workstations, a SuperMicro 4x H100 NVL, and a GH200.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
NVIDIA Releases 'PAIR' Beta for Local AI Inference Integration
Stanford and NVIDIA Release 'Contrastive Language Model' CLM with 9x Faster Inference