Video Processing Cost Optimization

The Cost Equation

Video processing cost per job:

Cost = (GPU hours × GPU rate) + (Storage GB × storage rate) + (Egress GB × egress rate)

Example:
GPU: 5 min × $1.00/hr ÷ 60 = $0.083
Storage: 500MB × $0.023/GB = $0.012
Egress: 1GB × $0.09/GB = $0.09
Total: $0.19 per job

GPU is usually 40-60% of cost. Focus there first.

GPU Utilization

The Utilization Problem

GPU instances cost the same whether busy or idle:

• 50% utilization = 2x your actual cost
• 25% utilization = 4x your actual cost

Target: 70%+

Leave headroom for bursts, but minimize idle time.

Strategies

• Autoscaling: Scale down when queue is empty
• Serverless: Pay only for actual processing (Modal, etc.)
• Batch processing: Accumulate jobs, process in batches
• Warm pools: Keep minimal instances for cold start, scale up for load

Spot Instances

60-90% Savings

Spot/preemptible instances cost much less but can be interrupted:

• AWS Spot: up to 90% off
• GCP Preemptible: 60-80% off
• Azure Spot: up to 90% off

When to Use

• Batch processing (not SLA-bound)
• Jobs that can checkpoint and resume
• Fault-tolerant workloads

When to Avoid

• Real-time or low-latency requirements
• Jobs without checkpointing
• SLA-bound customer workloads

Hybrid Approach

On-demand for SLA, spot for batch:

Priority queue → On-demand instances
Standard queue → Spot instances (with fallback)

Right-Sizing

Don't Over-Provision

Common mistake: using A100 when A10G suffices.

Workload	GPU	Cost/hr
720p enhancement	T4	$0.50
1080p enhancement	A10G	$1.00
4K enhancement	L40S	$2.50

Profile First

1. Measure VRAM usage
2. Measure processing time
3. Calculate cost per job
4. Find the sweet spot

Dynamic Selection

Choose GPU based on job:

if resolution <= 720p:
    use_gpu("t4")
elif resolution <= 1080p:
    use_gpu("a10g")
else:
    use_gpu("l40s")

Storage Optimization

Aggressive Cleanup

• Delete inputs immediately after processing
• Short output retention (24h-7d)
• Automatic lifecycle policies

Compression

• Store outputs compressed
• Use efficient codecs (HEVC over H.264)
• Adjust quality for use case

Tiered Storage

• Hot: S3 Standard (recent outputs)
• Warm: S3 IA (older outputs)
• Delete: after retention period

Efficiency Improvements

Batch Processing

Process multiple frames in one GPU call:

Batch size 1: 10 fps
Batch size 4: 35 fps (3.5x speedup)
Batch size 8: 50 fps (5x speedup)

Limited by VRAM — profile to find optimal batch size.

Model Optimization

• TensorRT for NVIDIA GPUs
• FP16 inference (vs FP32)
• Model pruning and quantization

Pipeline Parallelism

Overlap decode/process/encode:

While GPU processes frame N:
- CPU decodes frame N+1
- CPU encodes frame N-1

Keeps GPU fully utilized.