Skip to content

Storage options

NodeBoot DeviceStorageSize
clusterclaw (RPi5)MicroSDOS + Docker + OpenClaw + ThreadWeaver128GB
clustercrush (Orin Nano)MicroSDOS (JetPack)128GB
clustercrush (Orin Nano)NVMe M.2Models + llama.cpp256GB

Filesystem: ext4 (default for both Raspbian and JetPack)


ComponentSize
Raspbian Desktop (stripped + hardened)~3.4GB
Docker engine~500MB
OpenClaw container~1.2GB
ThreadWeaver container~800MB
Node.js 24~230MB
Blinkt! LED daemon~1MB
System overhead + logs~500MB
Total used~6.6GB
Free on 128GB card~115GB
  • Workload is mostly reads (Docker image layers, serving web UI)
  • No heavy writes (logs are capped, no database)
  • High-endurance microSD cards (Samsung PRO Endurance, SanDisk MAX Endurance) handle the write load
  • M.2 NVMe via HAT adds cost and case height for speed that isn’t noticeable in this use case
  • Heavy logging or data collection workloads
  • Running additional databases or storage-intensive services
  • Production deployments where SD card wear is a concern over years

clustercrush (Orin Nano) — Storage Breakdown

Section titled “clustercrush (Orin Nano) — Storage Breakdown”

The Orin Nano uses two storage devices:

ComponentSize
JetPack Desktop (stripped + hardened)~18GB
System utilities + security tools~1GB
System overhead + logs~500MB
Total used~20GB
Free on 128GB card~100GB
ComponentSize
llama.cpp build (CUDA)~500MB
Default model set:
Llama 3.2 3B Q4_K_M1.9GB
Llama 3.1 8B Q4_K_M4.7GB
Phi-3.5 Mini Q4_K_M2.3GB
Qwen 2.5 3B Q4_K_M2.0GB
Total used~11.4GB
Free on 256GB NVMe~230GB
NVMe SizeFitsUse Case
256GB~20 Q4 modelsStandard deployment
512GB~40+ Q4 modelsMulti-model testing, larger quantizations
1TBExtensive model libraryResearch, dataset storage
  • M-key, 2280 form factor
  • PCIe Gen3 x4 (~3.5 GB/s max)
  • Gen4 drives work but run at Gen3 speeds — no need to pay Gen4 premium

Only one model runs at a time on the 8GB Orin Nano (GPU VRAM limit). Use model-switch to swap between them.

ModelSize (Q4_K_M)ContextGen SpeedUse Case
Llama 3.2 1B0.7GB128K~80 t/sUltra-fast routing, classification
Llama 3.2 3B1.9GB128K~18 t/sPrimary agent model
Phi-3.5 Mini 3.8B2.3GB128K~17 t/sStrong reasoning
Qwen 2.5 3B2.0GB32K~18 t/sCode / structured output
Llama 3.1 8B4.7GB128K~10 t/sHighest quality; slower
Moondream21.9GB2K~15 t/sVision (lightweight)
Llama 3.2 Vision 11B6.2GB128K~7 t/sBest vision + language; tight fit
Terminal window
# On clustercrush:
sudo update-clustercrush.sh add-model https://huggingface.co/.../model.gguf
sudo model-switch new-model.gguf

  • Use high-endurance cards rated for continuous writes
  • Samsung PRO Endurance or SanDisk MAX Endurance recommended
  • Avoid consumer cards (Samsung EVO, SanDisk Ultra) — lower write endurance
  • A1/A2 application performance class preferred for random I/O
  • Any M.2 2280 NVMe drive works
  • Gen3 is sufficient (Orin slot is Gen3 x4)
  • Gen4 drives work but run at Gen3 speeds