Connect & discover
Connect heterogeneous GPU nodes and energy devices. Record hardware capabilities, connectivity and available resources.
GPU MANAGEMENT SOLUTIONSEvery GPU. Every watt.
Every GPU. Every watt.
One connected view.
Bring node health, workload dispatch and household energy into a single management experience. Explore how SunStack’s control platform is being developed for distributed compute.
SunStack control roomDistributed GPU operations
GPU nodes4 example nodes
Available3 ready for work
Energy hold1 waiting for surplus
GPU power675 W total
| Node / location | Hardware | Utilisation | Power | Status |
|---|---|---|---|---|
| Sydney · Node 01 | RTX 309024 GB VRAM | 78% | 225 W | Available |
| Newcastle · Node 02 | RTX 409024 GB VRAM | 64% | 270 W | Available |
| Wollongong · Node 03 | RTX 309024 GB VRAM | 42% | 180 W | Available |
| Dubbo · Node 04 | RTX 409024 GB VRAM | 0% | 0 W | Energy hold |
Example data and simulated controls. No live infrastructure is connected.
FROM DEVICE TO DECISIONFive responsibilities.
Five responsibilities.
One coordinated system.
A management approach that connects telemetry to decisions, while keeping device limits and household priorities visible.
Monitor & meter
Bring GPU utilisation, power, temperature, workload activity and dedicated energy metering into one operational view.
Understand the energy
Consider PV availability, battery state of charge, household load, tariffs and grid constraints before scheduling.
Predict performance & energy
Study how a job behaves on different GPU configurations and power limits to estimate runtime, energy and capacity needs.
Schedule & supervise
Match jobs to appropriate nodes, apply supported power controls, and respond to changing energy conditions or node failures.
The control room uses illustrative data. Predictive scheduling, automated power control and recovery are development capabilities, subject to hardware support and trial validation.
A PRACTICAL SYSTEM ARCHITECTURE
Local limits. Coordinated decisions.
The management approach connects observations from each property with workload planning and an operator’s view of the network.
Site telemetry
PV, battery, household and GPU metering.
Node runtime
Device health, capabilities and supported local controls.
The scheduler
Job matching, energy windows and recovery decisions.
Operator console
Fleet state, policies, workload status and measured outcomes.
PERFORMANCE & ENERGY PLANNINGBefore a job runs,
Before a job runs,
understand its options.
The prediction module is intended to estimate the consequences of choosing a GPU configuration and power setting. More GPUs do not automatically mean a better result.
Explore the evaluation approachDescribe the job and the hardware.
Task typeModel & memoryGPU generationGPU countPower limitEnergy windowDeadline
Compare useful outcomes.
Completion timeEnergy consumptionRequired capacityPower demandScaling efficiencyCost under agreed tariffs
Predictions need calibration against measured workloads. Multi-GPU estimates must reflect interconnect, task parallelism and system overhead.
CONTROL THAT RESPECTS THE SETTING
Respond to the conditions that matter.
Intended operating responses are subject to hardware support, workload behaviour and validation in trials.
| Condition | Management intent | What needs to remain visible |
|---|---|---|
| Household demand increases | Reduce new job admission or supported GPU power within the agreed policy. | Household load, GPU allocation and the reason for a change. |
| Battery approaches its reserve | Restrict battery-supported work and preserve the household reserve. | State of charge, reserve setting and eligible energy sources. |
| Temperature approaches a limit | Apply supported throttling, pause work or take the node out of service. | Temperature, effective power limit and recovery state. |
| A node becomes unreachable | Stop assigning new work and assess restart or migration options. | Job state, retry cost, output integrity and node connectivity. |
| A job needs more resources | Reassess placement and feasibility against model memory and runtime needs. | GPU capability, resource requirements and an explanation of the match. |
DEPLOYMENT & INTEGRATION
Start with visibility. Build towards control.
Discover the environment
Inventory GPUs, energy devices, metering, network boundaries and existing operating constraints.
Establish a baseline
Collect comparable telemetry and measure representative workloads before applying optimisation.
Run a controlled pilot
Introduce limited scheduling and supported controls, with clear fallbacks and reviewable results.
Expand with evidence
Use measured performance, household impact and operational outcomes to guide further deployment.
PLATFORM QUESTIONS
Designed for a mixed, changing fleet.
LET’S BUILD WHAT’S NEXT
Talk about GPU management