⚡ ScalePredict

Run a 2-min local benchmark → predict how long your AI job will take on cloud GPU.

No guessing. No wasted money.

---

The Problem

You have 1 million images to process with AI.
You open AWS and see:

T4   GPU  →  $0.52/hr
V100 GPU  →  $1.80/hr  
A100 GPU  →  $3.20/hr

You don't know which one to pick.
You don't know how many hours you'll need.
You guess. You pay. Sometimes you're wrong.

---

The Solution

python run_benchmark.py

2 minutes on your laptop. Then:

⚡ A100  →  0.4h   fastest
   V100  →  0.8h
   A10G  →  1.1h
   T4    →  2.3h

Look up the price yourself. Multiply. Done.

---

Quick Start

# Install
pip install -r requirements.txt

# Step 1 — measure your machine (2 min)
python run_benchmark.py

# Step 2 — open dashboard
streamlit run scalepredict_app.py

Opens at http://localhost:8501

---

Tested on Real Hardware

Machine	Type	Max Throughput	W Score
Lenovo L14	AMD Ryzen CPU	58 img/s	0.054
Xeon + Quadro M4000	Intel Xeon + GPU	639 img/s	0.730

CPU↔CPU correlation: r = 0.9969 — measured, not theoretical.

---

How It Works

run_benchmark.py
  → measures latency across batch sizes [1, 8, 32, 64, 128]
  → removes GPU warmup outliers automatically  
  → saves scalepredict_profile.json

scalepredict_app.py
  → reads your profile
  → applies k(t,d) scaling model
  → predicts runtime on T4 / V100 / A100 / A10G

The k(t,d) Model

k(t,d) = k₀ · e^(−αt) · (1 + β/d)

t  = batch size
d  = latency proxy (ms × 1000)
k₀ = architecture constant

This is the original formula behind the cross-architecture prediction.
Not a lookup table. Not a heuristic.

---

Files

ScalePredict/
├── run_benchmark.py      ← run this on your machine
├── scalepredict_app.py   ← Streamlit dashboard  
├── requirements.txt      ← dependencies
└── README.md

---

Requirements

Python 3.8+
torch >= 2.0
torchvision
psutil
streamlit
matplotlib
scipy
scikit-learn
numpy

---

Results Example

Running run_benchmark.py on Xeon + Quadro M4000:

batch=  1  →   5.9ms   170 img/s
batch=  8  →  17.0ms   470 img/s
batch= 32  →  53.8ms   594 img/s
batch= 64  → 104.4ms   613 img/s
batch=128  → 200.2ms   639 img/s

W score: 0.7295  ✅ Production ready

---

Roadmap

• CPU benchmark (Lenovo L14)
• GPU benchmark (Xeon + Quadro M4000)
• Streamlit dashboard
• Third machine validation
• arXiv preprint
• pip package

---

License

MIT — use freely.

---

Based on real measurements from 3 machines.
CPU↔CPU correlation r=0.9969.

Calculator (no install needed)

Just open the browser and enter your numbers: https://scalepredict.streamlit.app/calculator