← back to the showcase

What we have already sped up

Every line below is a real program that people wait for. We ran it on our test machine the normal way and our way, timed the whole job from start to finish both ways, and compared the results with the original. The numbers say how many times sooner the job finished: 2 times means the job took half as long. Each line links to a page with the full details: which program and version, what exactly we ran, and how you can check it yourself.

What you run What changes between runs How much sooner the job finished The results Details
Reports on tree models: which features matter, what-if questions (XGBoost, scikit-learn, CatBoost) the question you ask the model 2 to 10 times, by report (6.75×, 4.26×, 3.24×, 9.98×, 5.08×, 2.24×, 1.92×) every number identical Tree-model reports; LightGBM, measured separately
Strategy backtests and parameter searches (vn.py, backtrader, freqtrade, TA-Lib) the parameters you try 2 to 16 times, by case (vn.py 16.3×; backtrader 2.24–5.23×; freqtrade 5.91–11.02×; one TA-Lib indicator 7.11×) the same answers Strategy grids
Option pricing with Monte Carlo (QuantLib) the options you price 22.8 times on one pricing job identical to the last digit Strategy grids
Radiation shielding: trying different materials (OpenMC) the density of one material 3.55 times for ten scenarios (3.79 times against ten separate command-line runs) the same physics results Simulation families
Noise maps: trying different barriers (NoiseModelling) the barrier’s height, position or length, one at a time (three separate studies) 2.54, 2.37 and 2.39 times for ten scenarios each identical files Simulation families
Building models: checking clashes after moving parts (IfcClash) where the parts are 2.23 times for ten scenarios the same clash list Simulation families
Sunlight on terrain after editing the map (GRASS GIS r.sun) a patch of the terrain 2.38 times for ten scenarios identical files Simulation families
Genome scans after a change in the reference (Cas-OFFinder) a stretch of the reference sequence 5.77 times on chromosome 1; 6.07 times on chromosome 2 in a separate study; ten scenarios each the same hits Sequence search
Structures: many loads on one model (CalculiX) the loads 2.72 times for ten loads on a test model; 1.47 times on one real model from a public dataset identical files, apart from the run time stamps Load cases
Cleaning large 3D microscope images (SciPy) a small region of the image 6.96 times on a test image, 6.97 times on real microscopy data, ten edits each identical images Volume morphology

How to read the numbers

Your program

Tell us what you run: which program, how big the job is, how often you run it and how long you wait. The first check is free, and you pay only if your job got faster.

Source text: README.md

Telegram