GPU Memory Calculator & Optimization Guide

A practical reference for training PyTorch vision models without running out of GPU memory. Written for people who are new to GPU training (no systems background assumed).

---

What's included

Memory Calculator

Estimate peak VRAM before you run an experiment. 23 model presets including:

• Classification: MobileViT-XXS/S, EfficientNet-B1, ResNet-50, Swin-T/B, ConvNeXt-B, ViT-Base/Large
• Segmentation: SAM ViT-B/L/H, SAM2-Tiny/Small/Base+/Large, Mask2Former, SegFormer-B5
• Foundation models: DINOv2 ViT-S/B/L, CLIP ViT-B/L

The calculator uses architecture-aware spatial reduction factors (CNN=20, Hybrid=10, Transformer=50) to produce realistic estimates rather than naive worst-case numbers. Batch size configurable up to 2048.

Optimization Guide

A beginner-friendly reference covering:

• Start here — five changes that fix most GPU problems (AMP, DataLoader, no_grad, .item(), memory monitoring)
• What GPU memory actually is and why resolution is quadratic
• Mixed-precision training (AMP) and loss scaling
• Gradient accumulation, DataLoader tuning, CUDA prefetching
• Gradient checkpointing, frozen backbones, torch.compile()
• Classification vs. segmentation specifics (including SAM/SAM2)
• OOM troubleshooting with a prioritized fix order

---

Deploying to GitHub Pages

1. Fork or clone this repo
2. Go to Settings → Pages in your GitHub repo
3. Under Source, select main branch, / (root) folder
4. Click Save
5. Your site will be live at https://YOUR-USERNAME.github.io/YOUR-REPO-NAME/
6. Update the link in this README and in GPU_Memory_Calculator.html's nav GitHub link (id="gh-link")

No build step required. The entire site is a single GPU_Memory_Calculator.html file.

---

Files

GPU_Memory_Calculator.html  # The entire site — calculator + guide in one file
GPU Optimization Guide.pdf  # Formatted Word document version (downloadable)
README.md                   # This file

---

Formulas used

Activation memory (the dominant cost):

M_acts = B × (H×W ÷ sf) × C_avg × L × bytes × 1.5

Where:

• B = batch size
• H×W = input resolution
• sf = spatial reduction factor (CNN: 20, Hybrid: 10, Transformer: 50)
• C_avg = average channels per layer
• L = number of layers
• bytes = 2 for AMP/FP16, 4 for FP32
• 1.5 = overhead multiplier for autograd graph and intermediate buffers

The spatial factor corrects for the fact that CNNs progressively downsample (ResNet-50 goes 224→7px), so the average spatial footprint per layer is much smaller than the input resolution.

---

Notes on accuracy

The calculator produces estimates, not guarantees. Actual VRAM usage varies with:

• PyTorch version and CUDA version
• Specific model implementation details
• Whether you're in training vs. inference mode
• Fragmentation and caching behavior

Treat the output as a planning tool — add a 15–20% safety margin for production runs.

---

Contributing

Pull requests welcome. If you have calibrated measurements for a specific model that differ significantly from the estimates here, please open an issue with the model name, configuration, and measured peak VRAM.