GPU Infrastructure for Video Processing

GPU Options for Video AI

Modern video AI requires NVIDIA GPUs with sufficient VRAM:

GPU	VRAM	Best For	Cloud Hourly
T4	16 GB	Development, light inference	~$0.50
A10G	24 GB	Production inference	~$1.00
L4	24 GB	Efficient inference (newer)	~$0.80
L40S	48 GB	High-res, complex models	~$2.50
A100	40/80 GB	Training, large batches	~$3.00
H100	80 GB	Cutting-edge performance	~$4.00

VRAM Requirements

Video processing needs more VRAM than you'd expect:

• 720p video: 8-12 GB
• 1080p video: 12-16 GB
• 4K video: 24-32 GB
• 8K video: 48+ GB (tiled processing)

Add overhead for model weights (2-8 GB) and batching.

Cloud GPU Options

Major Cloud Providers

AWS

• G4dn (T4), G5 (A10G), P4d (A100), P5 (H100)
• Deep integration with S3, SQS
• Spot instances for 60-90% savings

GCP

• T4, L4, A100, H100
• Preemptible VMs for savings
• TPU alternative for some workloads

Azure

• NC series (T4), ND series (A100)
• Spot VMs for savings
• Strong enterprise relationships

Serverless GPU

Pay per second of actual compute:

• Modal: Python-native, fast cold starts
• RunPod: Container-based, flexible
• Banana: Simple deployment
• Replicate: Model-as-a-service

Serverless eliminates idle costs but adds cold start latency (5-30 seconds).

On-Premise Considerations

When On-Premise Makes Sense

• High utilization: >70% GPU usage
• Data sovereignty: Legal requirement for data location
• Predictable workload: Consistent, non-bursty demand
• GPU expertise: Team can manage infrastructure

Cost Comparison

A10G on-premise:
  Hardware: $10,000
  3-year TCO: ~$15,000 (with power, maintenance)
  Per-hour cost at 70% utilization: $0.81/hr

A10G in cloud:
  Per-hour cost: $1.00/hr
  Break-even: ~15,000 hours (2 years at 70%)

Hidden Costs

• Power and cooling
• Rack space
• Network infrastructure
• Maintenance and replacement
• Staff time

The Hybrid Approach

On-premise for baseline, cloud for bursts:

• Size on-premise for 50-60% of peak
• Burst to cloud for spikes
• Best of both worlds

Right-Sizing Your Infrastructure

Profiling Your Workload

Before choosing GPUs, understand your workload:

• Resolution distribution (720p, 1080p, 4K)
• Processing type (upscale, restore, denoise)
• Batch size (one video or many)
• Latency requirements (minutes OK or seconds?)

Benchmarking

Test on representative workload:

GPU      | 720p/s | 1080p/s | 4K/s  | $/1000 frames
T4       | 12     | 5       | 1.5   | $0.04
A10G     | 30     | 15      | 5     | $0.03
L40S     | 50     | 25      | 10    | $0.05

Common Mistakes

• Over-provisioning: A100 when A10G suffices
• Under-provisioning: T4 for 4K workloads
• Ignoring VRAM: Running out mid-job
• Ignoring cold start: User-facing latency suffers

BetterVideo's Infrastructure

We run on Modal's serverless GPU infrastructure:

Why Serverless

• Scale to zero: No idle costs
• Instant scaling: Handle bursts without provisioning
• No ops: No GPU maintenance
• Latest hardware: Access to A10G, L40S without purchase

Our GPU Selection

• A10G: Standard processing (up to 4K)
• L40S: High-resolution, complex models

Cold Start Optimization

• Models cached in container image
• Warm pool during peak hours
• Cold start under 5 seconds

You get GPU power without GPU infrastructure complexity.