Skip to main content
GPU Plexus

Flagship product

Swarm Manager

A single control plane for multi-tenant GPUs, from creating a container through to auditing what ran.

Platform targets

<60s
Container ready with SSH
<30s
GPU returned to the shared pool (Hibernate)

Core capabilities

Swarm Manager — run the whole GPU cluster from one control plane

Swarm Manager shares the GPUs in one machine across tenants, enforces real quotas, hibernates idle work to return resources, and restores MIG layouts automatically after a reboot.

Contact us →
  • MIG Partitioning

    Split a GPU into the MIG profiles the hardware supports so several jobs can share one card.

  • Smart Hibernate

    Pause idle containers, return the GPU to the shared pool and keep the data persistent.

  • Reboot Recovery

    The system is designed to restore the layout within 120 seconds of a reboot.

  • Same-Node Guarantee

    Jobs that use several GPUs are always placed on the same host.

  • One GPU per Container

    Reserve one specific GPU for one container.

The problems infrastructure teams meet every day — and how we handle them

Illustration of one GPU server with several teams connected to share its cards.
Illustration — one card shared by several teams, not a screenshot of the live system.

The usual way of using GPUs

  • A whole GPU is booked for one job, so memory sits idle and real utilization stays low.
  • There is no central quota, so teams compete for the same resources.
  • Software licensing for GPU management is a per-card annual cost paid abroad.

The GPU Plexus approach

  • Partition cards with MIG and hibernate idle work with Smart Hibernate to return GPUs to the shared pool.
  • Five-dimensional quotas (vCPU, RAM, ZFS Storage, GPU count, VRAM) with three permission levels.
  • Built by Thai engineers on the NVIDIA Driver standard, reducing reliance on foreign licensed software.

GPU utilization target

75–80%

Capacity increase target

30–50%

Target: from zero to an AI-ready environment in under a minute

No manual scripts — the system runs the steps in order.

  1. Step 1

    Choose an image and embed the SSH key

    The OS image is combined with the public key as soon as the create request arrives.

  2. Step 2

    Check five-dimensional quotas

    vCPU, RAM, ZFS Storage, GPU and VRAM are checked before allocation.

  3. Step 3

    Two-level smart node selection

    General jobs spread across CPU nodes; multi-GPU jobs are locked to one node.

  4. Step 4

    Create the ZFS dataset and attach MIG

    A volume is created with its real quota and linked to the MIG profile.

  5. Step 5

    Open Ingress and SSL automatically

    A port is allocated from the configured range and the domain is bound with its certificate.

  6. Step 6

    Ready to use

    The container is running and reachable over SSH.

Telemetry every 15 seconds

The system records resource usage every 15 seconds.

Illustration only — not live system data.

Three permission levels

  • Admin — Create and close accounts, set quotas and see every container.
  • Reseller — Share part of its own quota with customers and see only the work it manages.
  • User — Use containers within the assigned quota.

Five-dimensional quotas

  • vCPU
  • RAM
  • ZFS Storage
  • GPU count
  • VRAM

Audit Trail

A 365-day history of significant events.

Sample event names

  • container.created
  • user.resource_allocated
  • user.role_changed

Six core platform modules

Every module works from the same data, so what you see in the portal matches the real state of the cluster.

Illustration of blue and violet glass blocks clustered into layers.
Illustration — modules that make up one platform, not a screenshot of the live system.
  • Container

    Create, start, stop, hibernate and wake workloads, with health checks, environment, cron and live logs.

  • Image

    Assemble the OS image with the user’s SSH key, support private registries and .tar imports.

  • Volume & Storage

    A shared volume tree on ZFS with quotas, daily snapshots and backup and restore.

  • Port & Ingress

    Allocate ports from the configured range and attach domains with certificates automatically.

  • GPU & MIG

    Reserve one GPU per workload, apply the supported layout, and report drift after a reboot.

  • Fleet & Monitoring

    Register nodes, watch heartbeats, place work by remaining capacity and collect metrics every 15 seconds.

Engineering transparency

The design choices that make the system predictable and fixable in practice.

  • Intent separated from actual state

    The database holds what the system intends; actual state is read from the cluster and cached briefly to reduce read load.

  • Spec changes rebuild the container

    Changing a specification runs as a fresh container with rollback if any step fails.

  • Waking against real resources

    Waking a container matches a GPU set with enough VRAM, not merely a free host.

  • Drift is reported, not hidden

    When a GPU layout differs from the recorded one, the platform reports it for an operator to decide.

Three delivery models

Choose the model that fits how your organisation already runs infrastructure.

Illustration of a server room with several GPU servers mounted in racks.
Illustration — on-premise deployment, not a photograph of a real site.
  • Software License & Subscription

    Swarm Manager installed on your own infrastructure (self-hosted / on-premises), licensed per node or GPU per year.

  • Implementation & Support

    Design and installation of AI server infrastructure, software customisation to your requirements, and annual service-level support.

  • GPU-as-a-Service

    Rent containers with GPU and VRAM by quota specification and usage time — suited to demand that changes.

Built for teams that need the most out of every GPU

  • 01

    Fine-grained GPU sharing

    Use MIG to share a card between jobs without booking the whole device.

  • 02

    Automatic resource reclaim

    Hibernate idle work and keep data persistent with ZFS snapshots.

  • 03

    Real quota control

    Five-dimensional quotas and three permission levels are checked before every container is created.

  • 04

    Runs inside your organisation

    Self-hosted licence or managed service, without locking you to a single provider.

  • 05

    Supported by a Thai team

    Built and supported by engineers in Thailand.

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.