Fig. 3: Environmental impact of training and inference.
From: Ecologically sustainable benchmarking of AI models for histopathology
Training and inference impact the environment on different scales. Due to the quantity of data needed for training and backpropagation being disabled for model inference, an inference run consumes much less energy than training a model. a, b show the energy consumption and CO2eq footprint for training (n = 1) and c, d show the mean energy consumption (n = 15) and CO2eq footprint for inference for each evaluated model, respectively. e shows the number of usage of a model in a certain country, before one temperature related excess death occurs because of the emitted CO2eq, calculated on the basis of the mortality cost of carbon18. Based on the number of usage, the number of positive predictions are shown in (f). Panels e, f were calculated based on the mean CO2eq of (d). AU arbitrary units, KTX kidney transplant.
