Zero-Code-Change ML Acceleration with NVIDIA cuDF & cuML
Accelerating tabular machine learning on Google Cloud Colab Enterprise. Predict NYC yellow taxi tip amounts across millions of trips at 13.1x faster speed without rewriting standard pandas or scikit-learn code.
Pipeline Benchmark: CPU vs. GPU
Execution time in seconds for each phase on NVIDIA L4 vs. n2-standard CPU.
Live NYC Taxi Tip Prediction Simulator
Simulate predictions from the three cross-validated regression models and the optimal weighted ensemble.
Code Execution Architecture
Explore each phase of the accelerated machine learning pipeline.
Parquet partitions read straight to GPU memory with column downcasting (float64->float32), slashing memory footprint by 50%.
import pandas as pd
df = pd.read_parquet(...)
Scikit-Learn algorithms automatically offloaded to NVIDIA GPUs for cross-validation without changing standard function signatures.
from sklearn.ensemble import ...
model.fit(X_train, y_train)
Line-by-line inspection identifies slow row-by-row CPU fallbacks (e.g. .apply()) to convert them into 100% vectorized GPU kernels.
df['time'] = pd.cut(df['hour'], ...)
# 100% GPU vectorized
Ready to run on your own GPU environment?
Clone the repo, run in Google Colab Enterprise, or benchmark locally with RAPIDS.