On-premises
Processing and storage run in the customer-controlled environment. Define server capacity, network access, data retention, certificate management, backups and operational ownership.
Edge
Place decoding and inference near video sources when bandwidth, latency or data locality requires it. Define the information sent upstream and the behaviour during interrupted connectivity.
Distributed and hybrid topologies
Combine local processing with central management or review according to the delivery design. Node roles, resource allocation, synchronisation and failure behaviour must be specified; offline operation is not implied by the word edge.
Cloud components
Cloud placement is considered where permitted by data-handling rules, connectivity and source access. Estimate sustained media transfer and processing demand before selecting the topology.
Validate on target infrastructure
Use the actual models, source profiles and scenario settings for capacity tests. There is no universal cameras-per-GPU figure. Include decode load, resolution, frame rate, model input dimensions, concurrent graphs and evidence retention.
Record component placement and transfer paths
| Layer | Place explicitly | Check |
|---|---|---|
| Sources | Cameras, VMS and devices | Stream access and transport |
| Processing | Decoding, inference and scenarios | GPU profile, models and input data |
| Management | Registries, entity stores and APIs | Authorization, versions and task assignments |
| Data | Databases, MinIO and archive | Bandwidth, capacity and recovery |
| Operations | Metrics, logs and backups | Access from the administration network |
Estimate transfer of source video, model packages and results separately. Processing near the camera reduces upstream video traffic but still needs local capacity and delivery of model/scenario versions. For an interrupted connection, specify unfinished-job handling, data recovery and supported re-delivery; verify these in the actual deployment.