Table 1 Energy consumption and CO2eq emission for model training and inference

From: Ecologically sustainable benchmarking of AI models for histopathology

 

Training

Inference

Model

Energy (kWh)

CO2eq (kg)

Energy (Wh)

CO2eq/Slide (g)

TransMIL

11.263

4.065

0.128

0.046

CLAM

11.713

4.228

0.132

0.048

InceptionV3

10.584

3.821

0.201

0.073

ViT

23.873

8.618

0.170

0.065

Prov-GigaPath

42.625

15.388

0.63

0.229

  1. Energy consumption and CO2eq emission for model training and inference for the evaluated models TransMIL, CLAM, InceptionV3, ViT, and Prov-GigaPath. The CO2eq is calculated from the energy measured directly from low-level APIs multiplied by the local values for carbon intensity (Germany). For training, it was assumed that each model would be trained once in the time period, for 300 epochs, with a dataset with 1000 WSIs with 1000 patches, each with resolutions of 224px/256μm. CO2eq for inference was measured on one WSI with 1000 patches and of patch resolution of 224px/256μm.
Back to article page