Edge or Shared GPU for Cameras: The Economics of Computer Vision

Two ways to deploy vision on a production floor

Computer vision in a smaller factory usually follows one of two architectures. One places a compact edge computer beside each camera or line. The other sends several streams to a shared GPU server. Both can remain on premises: the question is where inference runs and what it costs to meet the required SLA.

Edge reduces network dependence and limits failure radius. A central server can use expensive acceleration more efficiently across multiple streams and simplifies model updates. The right choice depends on the cost of an accepted inspection, including downtime, wrong decisions and redundancy, not the sticker price of a box.

Measure workload before reading GPU specifications

OpenVINO's official `benchmark_app` distinguishes latency and throughput modes. The first minimizes time for one request; the second saturates a device with parallel work. NVIDIA DeepStream batches multiple video sources and recommends aligning batch size with source count because mismatches can add latency.

These tools provide a method, not a promise for a particular line. Inference-only benchmarks usually exclude video decode, ROI extraction, post-processing, event storage and PLC logic. Test the full pipeline at the real resolution, frame rate and lighting.

Record:

• camera count and the FPS actually required;
• maximum frame-to-signal latency;
• average and peak simultaneously active lines;
• encoded bitrate or frame size;
• acceptable downtime and the cost of a missed defect;
• required N+1 redundancy.

A model example with four lines

Assume four cameras, each requiring 10 analyzed FPS. The selected full pipeline achieves 55 FPS on a shared GPU server and 14 FPS on one compact edge device. These are model assumptions and must be replaced with measurements.

Edge option:

• four devices at 90,000 rubles each: 360,000;
• 50 W each, 200 W total;
• local event storage and four service points;
• one device failure affects one line.

Central option:

• a 360,000-ruble server plus 80,000 for networking: 440,000;
• 350 W under load;
• one update point, but one failure affects all four lines;
• true N+1 requires a second server, taking capital cost to 800,000 rubles before networking.

Without redundancy the central design is easier to operate, but its downtime radius is larger. With N+1, edge may cost less despite lower aggregate compute utilization.

The cost of traffic and cabling

An uncompressed 1920×1080 stream at 24 bits per pixel and 10 FPS is about 498 Mbps per camera before protocol overhead. Four cameras approach 2 Gbps. Production systems normally encode video or transmit only ROIs, but encoding adds latency and compute, while compression can hide small defects.

With edge inference, the network carries metadata, exception snapshots and selected frames. A central server needs stable transport for every stream. TCO therefore includes switches, fiber or copper, redundant links, installation, packet-loss diagnosis and OT segmentation.

An existing network is not free. Calculate the marginal upgrade caused by video and the engineering hours required to operate it.

Downtime dominates electricity

Energy rarely decides the architecture. At 10 rubles per kWh, the difference between 200 W and 350 W is about 13,140 rubles per year at continuous operation. One hour of line downtime may cost more.

Use a downtime-risk model:

`expected impact = unavailable hours × affected lines × cost per line-hour`.

If one edge box is unavailable eight hours per year and a line-hour costs 30,000 rubles, direct impact is 240,000 rubles. If an unprotected central server is down for four hours but affects four lines, impact is 480,000 rubles. These assumptions do not predict failure; they show why MTTR and blast radius belong beside CAPEX.

Some plants can continue with manual inspection. Then count extra labor, lower speed and the increased probability of missed defects rather than full downtime.

Cost per accepted inspection

Compare designs with:

`(depreciation + energy + network + operations + downtime + human review) / accepted inspections`.

Accepted means the system met latency, did not drop the frame and produced a process-usable result. High FPS is not valuable if the model creates an expensive queue of false alarms.

Include error costs. False rejects increase review and scrap; missed defects create claims and rework. Use the same model threshold on edge and server or the comparison describes different processes.

When edge wins

Edge is usually rational when:

• decisions are needed in tens of milliseconds near the machine;
• the network is unreliable or cannot carry video;
• lines must fail independently;
• data must remain within one production area;
• camera count is low and a central GPU would sit idle.

The tradeoff is more devices, updates and spare parts. A shared model registry, remote diagnostics and reproducible images are required or four inexpensive boxes become four different systems.

When a shared server wins

Centralization helps when cameras are intermittent, their peaks do not coincide, the model is heavy and the network already exists. Multi-stream batching can improve GPU utilization. Updates and monitoring stay in one place.

Redundancy must be explicit. “We will add the second server later” means accepting a single failure point today. That may be fine for non-critical analytics, but not for automated rejection across four production lines.

A hybrid is common: a simple model and emergency logic run at the edge, while heavy second-pass analysis and training run centrally. It creates two operating planes but can reduce traffic and preserve resilience.

A pilot before buying hardware

Collect 30–60 minutes of representative video from each line: normal production, defects, lighting changes, stoppages and peaks. Run the same pipeline on edge and server candidates. Measure end-to-end p95, dropped frames, power, utilization, bitrate and recovery time after a disconnect.

Then inject failures: remove the network, stop one inference process, deploy a model and roll it back. Economics without these tests counts hardware but ignores operations.

What management should ask for

Request two three-year estimates: per-line edge and a shared server with honest N+1. Both must use identical quality, latency and availability requirements, and report cost per accepted inspection.

The next step is a one-week shadow test on one line with p95 measurements and human-labeled errors. Scale only after that: cameras appreciate evidence, and budgets appreciate it even more.