The All Spark Cluster: Upgrading from 16 to 36 DGX Sparks
Kurcide · reddit · 2026-08-23
The author expanded their homelab DGX Spark cluster from 16 to 36 units, providing 4.6TB of unified memory, detailing network switches and cabling configurations.
Architecture & Usage
- Beyond single-point inference, the cluster is split into inference modules managed by Hermes and a custom memory sidecar, forming a persistent agent capability cluster.
- Expansion Goal: Dedicate 16 nodes to SOTA models (e.g., Kimi K3) while using remaining nodes for reranking, embeddings, and video/image/audio generation simultaneously.
Hardware Rationale
- Also deploying two RTX 6000 Pro systems to replace H100s/GH200s.
- Avoiding cooling/power issues of B200/B300; prioritizing full sovereignty without datacenter reliance.
- Believes Sparks offer the best value for scalable unified memory.
- Sparks + 6000 Pros offer flexibility, power optimization, and better liquidity for upgrades.
More from Embodied
- Realbotix: the Las Vegas company betting on robots that look very much like humans — kyliebytes · 2026-10-16
- Stanford's Fei-Fei Li team open-sources OpenWAM, a unified world-action model framework for robots — jiqizhixin · 2026-10-16
- Robot folding leaderboard: GPT-6 Astra leads at 1349 Elo, Muse Wonder trails at 1034 — qinzytech · 2026-10-16
- NVIDIA's TacitVLA: robots reason without reading their own thoughts, 3.4x lower latency — arankomatsuzaki · 2026-10-16
- SEAR: first large-scale controlled study of self-evolving robot agents, 7 LLMs, 330 tasks — xiye_nlp · 2026-10-16
- Podcast: subsea robotics startup Ulysses on operating 1,500m underwater and ocean carbon — blakemenezes · 2026-10-16