Current foundation doc uses references from a paper that only includes GPU; this is inaccurate and should be updated to include all energy use per https://sci.greensoftware.foundation
Comment:
GPU energy is probably one of the main contributors, but the CPU consumption of data loaders should also be included when discussing AI models. This can be very significant on certain types of models (training loops for VLMs that process video, for example, can take around 30-40% of the total batch compute just on data loading and decompression)
Current foundation doc uses references from a paper that only includes GPU; this is inaccurate and should be updated to include all energy use per https://sci.greensoftware.foundation
Comment:
GPU energy is probably one of the main contributors, but the CPU consumption of data loaders should also be included when discussing AI models. This can be very significant on certain types of models (training loops for VLMs that process video, for example, can take around 30-40% of the total batch compute just on data loading and decompression)