Deployment starts from your installed servers

Your institution racks the hardware, installs Ubuntu, and defines its network and storage environment. BareStax installs the control plane, joins and validates GPU nodes, configures BareStax Core, and keeps the cluster on a declared software state.

The same deployment model works on connected and fully air-gapped networks. No BareStax runtime component requires internet access.

The cluster starts from servers you have already racked

AreaSupported deployment
GPU nodesNVIDIA data center and workstation GPUs, including HGX systems from Hopper and Blackwell generations and servers with PCIe GPUs. Validated hardware classes are listed in each release note.
Cluster sizeSmall profile: 2–4 GPU nodes. Standard profile: 5–128 GPU nodes, up to 1,024 GPUs, with a separate head node.
NetworkInfiniBand and RoCE are supported, with NVLink and NVSwitch where present. These fabrics are optional; Ethernet-only servers can join.
Operating systemUbuntu 24.04 LTS on amd64. BareStax preserves the institution's operating system installation and does not erase the disk.
Head nodeA physical server without GPUs or a VM on the institution's existing virtualization platform. It hosts the control plane and cluster-local services.
Login nodeThe researcher entry point for SSH, the portal, and job submission.
StorageNFS and S3 targets. BareStax configures clients, mount protection, health, and impact; it does not install the storage server.
BMCRedfish-compatible BMCs. Inventory, sensors, event logs, and power state are visible; the Security approval policy governs power changes.

The full GPU is assigned directly to a job. BareStax does not use MIG or vGPU, and it does not place Kubernetes or a virtualization layer between BareStax Core and the GPU.

Installation puts the head node before compute capacity

Install barestax-server on the head node and create the cluster identity. The setup records the endpoint and writes the key backup to the specified file.

Connect LDAP, Active Directory, OIDC, SAML, or local identity, then configure built-in TOTP where required. Define cluster networking, NFS or S3 targets, the local package path, queues, tenants, quotas, and fairshare policy.

Test the alert chain before admitting compute capacity. The test confirms that an alert reaches the on-call recipient and that its code returns through the response path.

barestax-server init --cluster-name <name> --endpoint <fqdn> --key-backup <file>
barestaxctl alert test
barestaxctl doctor

Before a GPU node joins, BareStax checks firmware minimums, disk, Secure Boot, and connectivity. The server then registers with a token and waits for approval.

After approval, BareStax inventories the hardware, chooses the profile, installs the compatible stack, and runs GPU, NVLink, and InfiniBand checks plus two-node NCCL validation. Only a passing node becomes ready in BareStax Core.

Air-gapped releases cross a controlled boundary

On a connected network, barestax-bundler collects required packages, repository metadata, signatures, scan information, and compatibility data. It produces a signed, content-addressed release bundle and verifies it before transfer.

Approved portable media or a data diode carries the bundle into the isolated network. Import checks the signature, integrity, compatibility, and scan report. Four-eyes approval then promotes the bundle to the cluster-local repository.

Rollout starts with one canary node. Running jobs are preserved as BareStax continues through the cluster within the configured disruption budget.

barestax-bundler export
barestax-bundler verify
barestaxctl release import <bundle>
barestax-server upgrade --offline
01Bundler collects andsigns02Transfer by approvedmedia or data diode03Verify signature,integrity,compatibility, andscan report04Four-eyes approval05Canary node first06Roll out across thecluster within thedisruption budgetInternet sideAir-gap boundaryInternet side01Bundler collects and signsAir-gap boundary02Transfer by approved media ordata diode03Verify signature, integrity,compatibility, and scan report04Four-eyes approval05Canary node first06Roll out across the clusterwithin the disruption budget

Updates preserve compatibility and running work

BareStax distributes the kernel, DOCA-OFED, NVIDIA driver, CUDA libraries, Fabric Manager, DCGM, NCCL, and Apptainer only as tested sets. Timestamped repository snapshots and recorded key fingerprints fix the contents of each release.

The node rollout uses canary order and the configured disruption budget. The control plane upgrades through the same signed release process.

Each minor release line receives support for 24 months from its initial general release. Security fixes continue on that supported line, so an air-gapped institution can import the relevant bundle through the same controlled path.

Licensing and support remain available offline

The BareStax license works without a connection to a license server. Runtime operation does not send usage, identity, or cluster data outside the institution.

Support levels define response and resolution times by severity in the contract. Optional source code escrow is available. Pricing is provided after a technical discussion of cluster scope and support requirements.

The deployment review starts with node classes and transfer method

Walk a BareStax engineer through your node classes, fabric, storage targets, identity source, transfer method, and disruption policy.