Assess multi-stream vision platforms using actual video and business workloads. Advertised processor capacity, an isolated model benchmark and a single-stream demonstration describe only parts of their respective conditions. Procurement should place the intended streams, image modes, models, recording and interaction tasks in one verification configuration.
Break the workload into observable stages
| Stage | Main inputs | Information to record |
|---|---|---|
| Video input and decoding | Streams, codec, resolution, bitrate and timing | Actual reception, valid frames and decode errors |
| Preprocessing and transfer | Scaling, conversion, memory layout and copying | Processing time, memory and queue waiting |
| Model inference | Model version, input shape, numerical format and backend | Inference latency, throughput and resource use |
| Postprocessing and business rules | Detections, tracking, fusion and interfaces | Result generation and business acknowledgement time |
| Storage, display and operations | Images, evidence, recording, UI and logs | Concurrent effects, capacity and error states |
Transfers between stages also consume resources. Accelerating one stage may move the bottleneck to decoding, memory, storage or a business interface. Inference time alone does not define the device configuration.
Define throughput and latency separately
Throughput describes completed tasks in a specified interval. Latency describes elapsed time between defined start and end points for a task. Record per-stream valid outputs alongside aggregate results so a total does not hide a stream that waits indefinitely.
Official OpenVINO benchmark documentation reports inference latency and throughput separately and notes the effect of the test environment. This is a measurement reference, not evidence of arbitrary-device or complete-system performance.
Fix conditions before comparing platforms
Record the model and input, decoding and inference backends, OS and driver, concurrent tasks, power mode and cooling. Distinguish warm-up, steady operation and recovery. Explain the comparison again when models, dimensions, numerical formats or storage policies change.
Procurement and pilot verification
Begin with representative video and target operating modes, then increase streams and concurrent functions. Specify each stream's input and output, end-to-end latency, dropped-frame or intentional sampling policy, resources and duration. Without a target-scale test, mark capacity as pending verification rather than promising a fixed stream count.
Related services: edge AI deployment and optimization, embedded image acquisition and project enquiries.