Running work
Containers and storage
What your image needs, the resources each GPU brings, and the four paths mounted into every job.
The image
Any OCI image for Linux x86_64 with a CUDA 12.8 or newer userspace: 12.8 is the first CUDA release that builds for Blackwell. The host runs NVIDIA driver 580 and CUDA 13.0, so CUDA 12.8, 12.9 and 13.x images all work. We pull from public registries, and from private ones whose credentials you add in the console under Registries.
Images are cached per node for 14 days, so the second run of the same tag skips the pull. Pull time is part of the heat stage and is never billed.
Resources per GPU
Each GPU brings 14 CPU cores, 250 GB of memory and 3.75 TB of scratch. Asking for 8 GPUs gives you the whole node: 112 cores, 2 TB of memory, 30 TB of scratch and the full NVLink 5 fabric. The container runs as the user in the image and sees only its own GPUs.
Mounts
| Path | What it is | Lifetime |
|---|---|---|
/scratch | Local NVMe on the node. Fastest place to write. | Wiped when the job ends |
/shared | Your team's volume, the same in every job. | Kept while your account is active, plus 30 days |
/cache | Read only cache of popular open model weights, already on the node. | Managed by us |
/secrets | Files you registered in the console, read only. | Per job |
Shared storage
Each team gets 2 TB on /shared at no charge. Beyond that it is 0.06 USD per GB per month, measured daily. Write checkpoints and results here. It is replicated inside the forge and snapshotted nightly.
The model cache
/cache holds widely used open weights, laid out like a Hugging Face hub cache. Point HF_HUB_CACHE=/cache/hf at it and common models load without a download. The list is in the console.
Getting data in and out
forge cp ./dataset.tar f1:/shared/data/ # upload
forge cp f1:/shared/runs/run-14/ ./run-14/ # downloadIngress is free. Egress is free up to 5 TB a month per team, then 0.015 USD per GB. A job can also pull from your own bucket at start; outbound traffic from jobs counts as egress.