GPU Snapshot Timeline Explained

GPU Snapshot Timeline Explained | Generated by AI

Home 2026.06

Here’s the timeline from the forensic evidence:

March 16, 2026 — AMD built the snapshot image:

rocm:latest Docker image built with buildkit (26GB setup.sh = ROCm install layer)
droplet-agent purged (standard image-building cleanup)

May 20, 2026 04:39 — Your Droplet was created from that snapshot:

Cloud-init ran, re-installed droplet-agent
Host already had ROCm 7.2.0 at /opt/rocm-7.2.0/ (22G, baked into the image)

May 20, 2026 04:40 — Container rocm started:

Ran Jupyter Lab with ROCm GPU access
Exited ~May 21 (11 days ago)

Answer to your question: When you created the GPU Droplet, you chose an image like Ubuntu 7.2-gpu-mi300x1-192gb-devcloud-atl1. That image had both:

ROCm 7.2.0 installed on the host (/opt/rocm-7.2.0/) — 22G
rocm:latest Docker image pre-loaded (another 36G with ROCm inside) — with Jupyter Lab pre-configured

You didn’t install anything manually. AMD Dev Cloud baked it all into the snapshot. The Docker image is just a containerized duplicate of what’s already on the host — meant for running Jupyter Lab easily.

Bottom line: The 59G in containerd is entirely from the AMD Dev Cloud pre-loaded snapshot. The rocm container exited 11 days ago and you’re not using it. The host ROCm at /opt/rocm-7.2.0/ (22G) is what actually powers GPU workloads — the Docker image is redundant. Safe to nuke the Docker artifacts. Want me to clean them up?

Back Donate