Skip to main content

Shared Filesystems

Training large neural networks and fine-tuning models across multiple GPU nodes requires access to a unified storage repository. Having to duplicate a 200 GB dataset onto eight separate instances wastes time, disk space, and bandwidth.

NovaPod Shared Filesystems provide distributed network storage that mounts simultaneously to any number of your GPU instances.

Technical Specifications​

  • Throughput: Up to 25 Gb/s sequential read/write performance.
  • Protocol: POSIX-compliant distributed filesystem (NFSv4 / CephFS backed).
  • Multi-Attach (Read-Write-Many): Mount the exact same volume to 1, 10, or 100 GPU instances concurrently.
  • Persistence: Storage outlives any single instance. You can terminate all compute instances while your models, training logs, and checkpoints remain safely preserved.

Creating & Mounting a Shared Volume​

Via CLI​

# Create a 500GB shared filesystem volume in US-East
np storage create --name model-weights-vol --size 500GB --region us-east

# Launch Instance 1 mounting the shared volume to /data
np launch \
  --gpu H100_SXM \
  --storage-volume model-weights-vol:/data \
  --name trainer-node-01

# Launch Instance 2 mounting the same volume
np launch \
  --gpu H100_SXM \
  --storage-volume model-weights-vol:/data \
  --name trainer-node-02

Inside both instances, /data is the identical directory. Any checkpoint written by Node 1 is instantly readable by Node 2.

Pricing​

  • $0.00014 per GB per hour (~$0.10/GB/month).
  • Zero ingress or egress fees for internal data transfers between instances and storage volumes within the same region.