Compute Unit

Compute Units (CUs) are the standard measurement of processing consumption within EyePop.ai.

Whenever EyePop analyzes an image, video, livestream, or performs AI inference, Compute Units are consumed based on the amount of work performed.

Why Compute Units Exist

Different AI workloads require different amounts of compute. For example:

• Detecting a single object in an image requires less compute than analyzing a high-resolution video.

• Running multiple models in sequence requires more compute than running a single model.

• Processing a one-hour recording requires more compute than processing a single frame.

Compute Units provide a consistent way to measure usage across all EyePop products and capabilities.

What Affects Compute Unit Consumption?

Several factors influence Compute Unit usage:

Media Type

• Images

• Video files

• Live streams

Processing Duration

Longer videos and streams generally consume more Compute Units than shorter inputs.

Model Complexity

Advanced workflows involving multiple detections, tracking, OCR, or Vision Language Models (VLMs) may consume more Compute Units than simpler pipelines.

Resolution and Frame Rate

Higher resolutions and higher frame rates require more processing and may increase consumption.

Workflow Configuration

A pipeline that performs multiple steps such as:

Vehicle Detection → License Plate Detection → OCR

will consume more Compute Units than a single-stage detection workflow.

Monitoring Usage

Compute Unit consumption can be monitored through the EyePop dashboard.

Usage reporting helps teams:

• Track consumption trends

• Estimate future costs

• Optimize workflows

• Understand which applications consume the most resources

Billing

Most EyePop plans include a monthly allocation of Compute Units.

If usage exceeds the included allocation, additional Compute Units may be billed according to your plan.

For current pricing and included allocations, see the Billing & Pricing documentation.

Best Practices

To maximize Compute Unit efficiency:

• Process only the frames required for your use case.

• Use object detection to filter content before running more expensive analysis.

• Select the smallest model capable of solving the problem.

• Leverage tracking to avoid reprocessing the same objects repeatedly.

• Test workflows using representative data before deploying at scale.