Cloud GPU inference
Accelerates selected model processing and supports concurrent video analytics workloads.
SYSTEM ARCHITECTURE
Six layers. One connected pipeline.
Separate video ingestion, AI processing and reporting so each component can be evaluated and scaled around actual demand.
01 / PROCESS LAYER
Camera ingestion
Connect compatible video streams through authorised network paths.
02 / PROCESS LAYER
Stream processing
Decode video and select frames for model analysis.
03 / PROCESS LAYER
GPU inference
Run computer vision models and tracking components.
04 / PROCESS LAYER
Event logic
Apply counting lines, zones and configured conditions.
05 / PROCESS LAYER
Data management
Store approved events, aggregates and selected media.
06 / PROCESS LAYER
Dashboards and alerts
Present reporting, events and system status to authorised users.
VISION INFERENCE
OBJECT DETECTIONEVENT PROCESSINGAI / CLOUD GPUAI COMPUTE
Choose compute for your actual workload.
Cloud GPUs can accelerate person detection and other computer vision inference tasks. CPU and decoder capacity are equally important for maintaining the overall video pipeline.
- Plan for camera count, resolution and model complexity.
- Distinguish source frame rate from analysed inference FPS.
- Benchmark throughput, model quality and alert latency together.
- Monitor GPU memory, CPU load and processing backlogs.
CLOUD COMPONENTS
Infrastructure that supports reliable insights.
Video analytics needs more than a model endpoint. The surrounding compute, networking and storage make results usable.
CPU and application compute
Handles video ingestion, APIs, decoding where applicable and background processing.
Object and structured storage
Stores approved event clips, aggregates, system records and application data.
Private networking
Controls how streams, processing services and dashboards communicate.
Access and security
Supports permission controls, encryption, monitoring and responsible retention.
Operational monitoring
Tracks stream health, utilisation, inference performance and capacity trends.
OBSERVABILITY
Measure what the pipeline actually delivers.
The following should be captured from deployed workloads and labelled clearly by environment.
Processing FPS
Analysed frames per second, per stream and overall.
Model quality
Task-specific precision, recall and counting error.
Alert latency
Event-to-notification delay, including p95 where relevant.
Storage consumption
Retained video, event data, archives and backups.
| Metric | How to measure | Why it matters |
|---|---|---|
| Processing FPS | Count frames completed per second | Determines sustained pipeline capacity |
| Counting error | Compare estimated and verified crossings | Tests footfall usefulness |
| Alert latency | Event timestamp to user notification | Shows responsiveness for defined events |
| Storage consumption | Measure retained bytes by data type | Informs retention and cost planning |
| Stream availability | Monitor camera connectivity and interruptions | Exposes data gaps |
These are evaluation metrics, not claims of achieved SiliconCloud performance.
PRIVACY BY DESIGN
Responsible video analytics is part of the architecture.
Retail footage can contain identifiable individuals. Design choices should address the purpose of processing, approved access, data transfer, retention and human review before a deployment goes live.
- Collect only the data needed for defined analytics.
- Restrict camera credentials and dashboard permissions.
- Specify encryption, retention and deletion controls.
- Use human review for alerts that could affect people.
ACCESS CONTROL
DATA RETENTIONSECURE STREAMSHUMAN REVIEWGood to know
Frequently asked questions
Clear answers before you plan your video analytics project.
Does each camera need a dedicated GPU?
No. One GPU may process several streams, but capacity must be established through benchmarks of the real pipeline.
How much storage is needed?
It depends on bitrate, number of cameras, continuous versus event-only recording, retention and backup policies.
What is processing FPS?
The number of video frames the analytics pipeline processes each second. It may differ from the camera source frame rate.
What should count as model accuracy?
There is no one universal metric. Use precision, recall and task-specific measures such as counting error on representative labelled footage.
Can the infrastructure scale?
The architecture can be designed to add processing resources, but actual scaling behaviour must be tested with network, decode, inference and storage demands.