Inference, dispatch & metering
The existing project showcases streaming text and image-enabled inference, node dispatch, and energy and usage views.
RESEARCH & DEVELOPMENTBig questions.
Big questions.
Real-world purpose.
Exploring what happens when distributed AI compute meets household energy. Our research connects systems engineering, AI efficiency and the realities of Australian homes.
Scheduling, efficiency & trust
Energy-aware optimisation and stronger execution assurances form the next development priorities.
Evidence from real homes
Staged household deployments are intended to test practical performance and the participation model.
THE RESEARCH AGENDAFour connected
Four connected
lines of enquiry.
These are research directions and evaluation plans. Commercial performance claims will depend on measured results from deployment.
Good research starts with the right collaborators.
We welcome conversations with universities, energy researchers, GPU systems teams and industry trial partners.
FROM HYPOTHESIS TO EVIDENCE
Test the pieces. Then test the system.
A staged methodology helps separate a useful idea from a dependable deployment.
Controlled benchmarks
Measure GPU workloads on known hardware, model versions and power configurations.
Historical replay
Replay energy and demand traces to compare schedulers under the same conditions.
Hardware in the loop
Connect the decision logic to physical devices and measure response, limits and failures.
Staged household trials
Evaluate practical behaviour, installation constraints and participant experience in real environments.
MEASUREMENT FRAMEWORK
Claims should have a comparison.
These proposed measures provide a basis for evaluating the work. No numerical results or publications are implied.
| Research question | Comparison | Evidence to collect |
|---|---|---|
| Does energy-aware dispatch improve outcomes? | Identical jobs and energy traces under energy-blind and energy-aware scheduling. | Completed work, deadlines, energy per job, imports and reserve violations. |
| Can power control preserve useful performance? | The same model and task quality across supported GPU power limits. | Runtime, throughput, GPU and whole-node energy, temperature and output quality. |
| How resilient is the network? | Controlled loss of a node, connection or eligible energy window. | Recovery time, failed work, restart overhead, migration behaviour and output integrity. |
| What changes for the household? | Comparable household profiles with and without the proposed GPU policy. | Load impact, reserve behaviour, comfort constraints and clearly separated GPU consumption. |
| What assurance does execution provide? | Defined adversarial and fault scenarios with explicit trust boundaries. | Isolation results, attestation coverage, verification cost and incident recovery. |
RESEARCH THAT CAN BE EXAMINEDMake the method
Make the method
as clear as the result.
Hardware diversity, weather and household behaviour all affect the outcome. Evaluation needs to preserve that context.
Comparable conditions
Keep workload, model quality and execution settings aligned when comparing performance.
Energy boundaries
Distinguish GPU power from whole-node and household energy, and record where measurement takes place.
Uncertainty & failures
Report variability and unsuccessful runs alongside successful averages.
Reproducible artefacts
Plan versioned configurations, benchmark procedures and appropriately shareable evaluation data.
INTENDED RESEARCH OUTPUTS
Useful knowledge, beyond the prototype.
Performance profiles
Document how different GPU and power configurations affect useful work.
Planned outputScheduling methods
Evaluate decision strategies that connect workload needs with energy conditions.
Planned outputAssurance mechanisms
Study verifiable execution, isolation boundaries and failure handling.
Planned outputTrial evidence
Report what staged deployments reveal about real household operation.
Planned outputLET’S BUILD WHAT’S NEXT
Discuss a collaboration