toolbox
Home / AI, Data & Technology / AI Video Pipeline Throughput Estimator

AI Video Pipeline Throughput Estimator

This estimator calculates how many video jobs an AI processing pipeline can complete over a selected time window. It combines worker count, average processing time, parallel jobs per worker, utilization, and retry overhead to produce both theoretical and practical throughput.

Use it when sizing GPU workers, planning batch deadlines, or checking whether a rendering, transcription, enhancement, moderation, or generation workflow can meet demand. The practical estimate accounts for idle time and repeated work, so it is more useful for capacity planning than a simple jobs-per-hour figure.

Pipeline assumptions

min
hr
%
%
Result
—
Practical completed jobs
Theoretical completed jobs—
Practical jobs per hour—
Available worker-minutes—
Estimated retry workload—

1. Enter worker capacity
Enter the number of active workers and how many jobs each worker can process at the same time.

2. Set processing duration
Use the end-to-end average minutes per video job, including model inference and required post-processing.

3. Choose the operating window
Enter the total number of hours available for the batch or planning period.

4. Apply utilization
Reduce theoretical capacity for setup, queue gaps, maintenance, and other idle time.

5. Account for retries
Enter the share of attempted jobs expected to require reprocessing, then review practical throughput.

Theoretical jobs = Workers × Parallel jobs per worker × Operating hours × 60 ÷ Minutes per job
Practical jobs = Theoretical jobs × Utilization rate × (1 − Retry rate)

Utilization and retry percentages are converted to decimals. The model assumes average processing time and concurrency remain reasonably stable during the window.

What the result means

The main result estimates unique video jobs completed after expected idle capacity and retry workload are removed.

Actual throughput may differ when job lengths vary, queues are uneven, or workers have different hardware.

Given: 8 workers, 2 parallel jobs each, 12 minutes per job, 24 hours, 85% utilization, and 5% retries.

Calculation: Theoretical jobs = 8 × 2 × 24 × 60 ÷ 12 = 1,920. Productive attempts = 1,920 × 0.85 = 1,632. Practical completed jobs = 1,632 × 0.95 = 1,550.4.

Result: About 1,550 completed jobs in 24 hours, or 64.6 jobs per hour.

Should processing time include upload and post-processing?

Include every pipeline stage that occupies the constrained worker or delays its next job. Exclude steps that run independently without consuming that capacity.

What does utilization represent?

Utilization is the share of theoretical worker time expected to perform useful processing. Queue gaps, maintenance, startup time, and scheduling inefficiency lower it.

How should I estimate retries?

Use recent production logs when available. Count jobs that must be run again because of errors, quality failures, or interrupted processing.

Can workers with different speeds be combined?

A single average can be used for a rough estimate, but separate calculations by hardware class are more accurate. Add the practical throughput from each group.

Is this the same as maximum requests per second?

No. This tool estimates completed video jobs over time, while requests per second usually describes short API events and may ignore long-running processing.