Built for the messy middle of AI development
Everything between “idea” and “training run” lives on a node.
JupyterLab, zero setup
Prebuilt images ship with JupyterLab, tested CUDA versions and framework dependencies configured. Click Launch App and you're in a notebook — run nvidia-smi to see your GPU.
Prebuilt, Base OS or your container
Choose managed ML images, a minimal Ubuntu Base OS for full control, or any public/private Docker image. Build team-standard images with the TIR Image Builder and Container Registry.
Storage that fits the workload
Persistent workspace at /home/jovyan (30GB free, expandable to multi-TB), EOS/Disk Datasets for large data, Shared File System for team access, Parallel File System for high-throughput jobs.
SSH like a real machine
Toggle SSH access, attach one or more keys, open port 22 via a security group, and use your own tools — VS Code Remote, rsync, tmux. Start scripts automate environment setup on every boot.
Save Image
Snapshot your whole environment — packages, configs, scripts, frameworks — as a reusable image. Restore it after experiments or launch identical nodes for the whole team.
Monitoring, alerts, events
Live gauges for memory, workspace and ephemeral usage, CPU/RAM utilisation charts, threshold-based email alerts, and a full instance event log for troubleshooting.
Launch in five steps
The actual create-instance flow in the TIR console.
Pick an image
Pre-built (JupyterLab-ready), Base OS, or Custom. Enable "JupyterLab Supported" if your custom image was built for it.
Choose compute
GPU, CPU, Spot or Private Cluster. Pick the hardware plan (e.g. NVIDIA H200 SXM) and the vCPU/RAM configuration.
Pricing & storage
On-demand hourly or committed (1/3/6 months). Workspace starts at 30GB free; AES-256 storage encryption is a checkbox.
Network & security
Attach an SSH key (new keys download as .pem), pick a security group, optionally reserve a static IP or attach a VPC.
Launch
Review and launch — or do the whole thing via the REST API or Terraform.
Work, stop, resume
Compute bills only while Running. Stop overnight; your workspace is waiting in the morning.
Storage tiers on a node
Match the tier to the job — all attachable to a single instance.
| Tier | What it is | Persistence | Use it for |
|---|---|---|---|
| Workspace | Pre-attached disk at /home/jovyan — 30GB free, expandable (grow-only) | Persists | Code, checkpoints, working files |
| Datasets (EOS / Disk) | Object-storage or disk-backed data mounts; disk attaches to one instance at a time | Persists | Training corpora, large files kept out of workspace |
| Shared File System (SFS) | Concurrent read/write mounts (e.g. /my_sfs) across instances | Persists | Team-shared data and configs |
| Parallel File System (PFS) | Striped high-throughput storage for simultaneous access | Persists | HPC-style and multi-process data loading |
| Ephemeral | Temporary space at /home/user, fixed 50GB | Lost on restart | Scratch only — never production data |
Terminal-first? Same node, your tools
Attach a key and a security group with port 22 open, then:
# illustrative — use the connection details from your instance page
chmod 400 my-tir-key.pem
ssh -i my-tir-key.pem jovyan@<instance-ip>
# verify your GPU
nvidia-smi
# keep persistent work in the workspace mount
cd /home/jovyan && git clone https://github.com/your-org/your-model.gitFiles outside /home/jovyan may be lost on container restart — keep anything that matters in workspace, Datasets, SFS or PFS.
Outgrown a single node?
Graduate to managed fine-tuning, multi-node Slurm clusters, or production endpoints without moving clouds.
Frequently Asked Questions
Everything you need to know about AI Dev Nodes on TIR.
AI Dev Nodes
TIR nodes are container-based AI workspaces with JupyterLab, prebuilt ML images, dataset mounts and Save Image built in. If you want a plain virtual machine with full OS control, use Cloud Compute — same GPUs, different abstraction.
Plans span current NVIDIA hardware — including H200 SXM and H100 — plus CPU-only and spot capacity. Configurations pair each GPU with vCPUs and RAM (e.g. 1× H200 / 30 vCPU / 375GB RAM). See the rate card for what's live.
Compute billing stops with the node; workspace storage persists (storage charges may apply). Committed plans of 1, 3 or 6 months lower the hourly rate for always-on work.
Yes — save your configured instance as an image and share it via the TIR Container Registry, so everyone launches from the identical environment. Shared File System mounts give common data access across instances.
Yes, Spot is one of the instance types in the create flow — useful for interruptible experiments at lower cost. Checkpoint to persistent storage as a habit.
Stop configuring CUDA. Start building.
Get hours of GPU notebook time, on demand.