Skip to main content
GPU Plexus

Architecture

Architecture you can audit

Every layer has a clear responsibility, so the platform can be deployed on-premises, audited and recovered.

Five layers, each with one job

Each layer owns a single concern, which makes the system auditable, replaceable and recoverable without touching the whole.

  • Layer 1 — Presentation & API

    The web portal and REST API take user commands, with WebSocket for live logs.

    • Web Portal
    • REST API
    • JWT Cookie / OAuth2
  • Layer 2 — Access & Governance

    Middleware inspects headers, locks privileged operations and verifies permission on every request, alongside ES256 software licensing.

    • Middleware
    • 3 permission levels
    • ES256
  • Layer 3 — Service Layer

    Business logic for Container, Image, Volume, Port, GPU/MIG and Fleet modules is reached through one factory.

    • Container
    • Volume
    • GPU / MIG
  • Layer 4 — Orchestration & Node Agent

    Docker Swarm runs the workloads, while a privileged Node Agent handles host-level work: ZFS, snapshots, NFS, nginx and GPU probing.

    • Docker Swarm
    • Node Agent
    • GPU Probe
  • Layer 5 — Data & Storage

    PostgreSQL holds intended state, ClickHouse holds metrics and audit, and ZFS holds container data persistently.

    • PostgreSQL
    • ClickHouse
    • ZFS

Core technology

  • Docker Swarm

    Runs the cluster and places work on nodes by remaining capacity.

  • PostgreSQL

    The primary store for intended state and permissions.

  • ClickHouse

    The analytical store for metrics and audit.

  • ZFS

    Dataset-based storage with daily snapshots.

  • NFS

    Shares volumes between containers in the cluster.

  • ES256 Licence

    A digital signature bound to the machine fingerprint, working without an internet connection.

  • OAuth2 / SSO

    Authentication through the identity provider the organisation already uses.

  • NVIDIA MIG

    Partitions a GPU into the profiles the hardware supports.

Data and storage

User data has to outlive a hibernated or rebuilt container.

  • A shared volume tree on ZFS, where each dataset carries its own quota.

  • Daily snapshots with a defined retention window.

  • Backup and restore per dataset, including cloning into a new dataset.

  • Data stays in place when a container hibernates and returns when it wakes.

* A snapshot is not a substitute for off-system backup — plan additional backups to your own policy.

Security and governance

  • Per-user networking

    Each user has its own overlay network, limiting visibility between workloads.

  • Scoped tokens

    Personal tokens name the methods they may call and are rate limited.

  • Verifiable licensing

    The ES256 signature is bound to the machine and can be checked without staying online.

  • A full audit trail

    A 365-day trail stored in a database that cannot be edited after the fact.

  • Warnings before the limit

    Alerts at 80%, 90% and 95% of quota, with workloads suspended when they exceed entitlement.

Limits we state plainly

The current version draws its scope clearly so that what we deliver is predictable.

  • No interactive exec in the browser

    Working inside a container is done over SSH, as designed.

  • No built-in billing

    Charging for usage is a process outside the platform.

  • MIG layout changes need a helper

    They run through a one-shot privileged helper, not a normal portal command.

Ready to design AI infrastructure that fits your organisation?

Tell our engineers about your challenge and we will help assess the right approach and cluster size.