gpu-cost-optimizer

Optimize GPU compute costs in diffusion pipelines by up to 70%.

Measures workflow compute time on cloud GPUs and adjusts scheduler steps, resolution, and batch size to minimize cost while maintaining acceptable quality.

Installation

pip install gpu-cost-optimizer

Or from source:

git clone https://github.com/belentani7/gpu-cost-optimizer.git
cd gpu-cost-optimizer
pip install -e ".[dev]"

Quick Start

# Profile a workflow on a specific GPU
gpuopt profile workflow.json --gpu a100

# Optimize to meet a target cost
gpuopt optimize workflow.json --target-cost 0.50

# Compare across GPUs
gpuopt compare workflow.json --gpus a100,h100,l40s

# Generate a report
gpuopt report profile.json --output report.md

Workflow JSON Format

{
  "name": "character_generation",
  "pipeline": "stable-diffusion",
  "model": "sd-xl-base",
  "scheduler": {
    "type": "euler_a",
    "steps": 50
  },
  "resolution": [1024, 1024],
  "batch_size": 1,
  "guidance_scale": 7.5,
  "safety_checker": true,
  "vae_tiling": false
}

Optimization Strategies

Strategy	Typical Savings	Quality Impact
Reduce scheduler steps	40-60%	Minimal with turbo schedulers
Lower resolution	20-50%	Depends on target use
Optimize batch size	10-30%	None
Turbo scheduler swap	30-50%	Minor
Model quantization (FP16)	25-40%	Negligible

GPU Pricing Reference

Realistic on-demand pricing from major cloud providers (USD/hour):

GPU	VRAM	Typical Price
NVIDIA A100 80GB	80 GB	$2.21
NVIDIA H100 80GB	80 GB	$3.70
NVIDIA L40S	48 GB	$1.86
RTX 4090	24 GB	$0.74
RTX 3090	24 GB	$0.44
A10G	24 GB	$0.80

License

MIT

This site is open source. Improve this page.