Fast execution can conceal incorrect computation.
Validate numerical output and retain timing statistics before presenting measured throughput.
← Selected work
Browser-based measurement
CanMyPCRunAI
Measuring browser GPU workloads with numerical validation, controlled execution, and clear limits on what the results mean.
CanMyPCRunAIBROWSER WORKLOADS
Conceptual architecture · no benchmark results shown
THE PROBLEM
What needed to change
A performance number is only useful when its meaning is clear. Browser GPU measurement must account for numerical correctness, timing differences, unsupported features, and interrupted runs.
MY CONTRIBUTION
From concept to implementation
My approach centers on trustworthy measurement: check the computation before interpreting its speed, make interruptions recoverable, and explain what a browser benchmark can and cannot establish.
ARCHITECTURE
How the pieces fit
- A capability scan identifies available browser GPU features.
- Bounded WebGPU/WGSL workloads execute with reference-output validation and timing statistics.
- React presents measured categories, guidance, and an explicit JSON export.
- Separate development tooling supports calibration without presenting an experimental score as a public conclusion.
ENGINEERING DECISIONS
Constraints and solutions
A browser test can be interrupted or lose its GPU device.
Use explicit lifecycle states, bounded deadlines, cancellation, and resource cleanup.
Users may interpret throughput as a guarantee that a particular AI model will run.
Keep capability guidance separate from measurement and explain that browser signals do not establish model fit or installed memory.
RESULTS
What the work delivers
- Launched a public beta with real browser GPU workloads and downloadable measured results.
- Implemented capability-based guidance without converting throughput into unsupported hardware rankings.
- Built separate calibration and experimental-scoring tools for further research.
The launch handoff records automated checks, production-bundle guards, a completed hosted benchmark, JSON export, and responsive checks. Those dated results do not establish performance across all devices or real-model inference compatibility.
ACCESSIBILITY
Where human evaluation matters
Controls include status announcements, cancellation, and focus handling. Results are explained in text. These implementation choices support usability; they are not a claim of comprehensive accessibility conformance.