import streamlit as st import pandas as pd import os, csv from huggingface_hub import hf_hub_download, HfApi import math HF_TOKEN = os.getenv('HUGGING_FACE_HUB_TOKEN') CACHED_FILE_PATH = hf_hub_download(repo_id="sasha/co2_submissions", filename="co2_emissions.csv", repo_type="dataset") api = HfApi() def write_to_csv(hardware, gpu_tdp, num_gpus, training_time, provider, carbon_intensity, dynamic_emissions, experimentation_time, experimental_emissions, pue, pue_emissions, embodied_type, embodied_emissions, model_info): st.session_state["is_shared"] = True with open(CACHED_FILE_PATH,'a', newline='') as f: writer = csv.writer(f) writer.writerow([hardware, gpu_tdp, num_gpus, training_time, provider, carbon_intensity, dynamic_emissions, experimentation_time, experimental_emissions, pue, pue_emissions, embodied_type, embodied_emissions, model_info]) api.upload_file( path_or_fileobj=CACHED_FILE_PATH, path_in_repo="co2_emissions.csv", repo_id="sasha/co2_submissions", repo_type="dataset", ) st.set_page_config( page_title="AI Carbon Calculator", layout="wide", ) tdp_url = "https://raw.githubusercontent.com/mlco2/impact/master/data/gpus.csv" compute_url = "https://raw.githubusercontent.com/mlco2/impact/master/data/impact.csv" electricity_url = "https://raw.githubusercontent.com/mlco2/impact/master/data/2021-10-27yearly_averages.csv" server_sheet_id = "1DqYgQnEDLQVQm5acMAhLgHLD8xXCG9BIrk-_Nv6jF3k" server_sheet_name = "Server%20Carbon%20Footprint" server_url = f"https://docs.google.com/spreadsheets/d/{server_sheet_id}/gviz/tq?tqx=out:csv&sheet={server_sheet_name}" embodied_gpu_sheet_name = "Scope%203%20Ratios" embodied_gpu_url = f"https://docs.google.com/spreadsheets/d/{server_sheet_id}/gviz/tq?tqx=out:csv&sheet={embodied_gpu_sheet_name}" TDP =pd.read_csv(tdp_url) instances = pd.read_csv(compute_url) providers = [p.upper() for p in instances['provider'].unique().tolist()] providers.append('Local/Private Infastructure') ### Default values hardware = "N/A" gpu_tdp = 0 num_gpus = 0 training_time = 0.0 provider = "N/A" carbon_intensity = 0.0 dynamic_emissions = 0.0 experimentation_time = 0.0 experimental_emissions = 0.0 pue = 1.0 pue_emissions = 0.0 embodied_type = 0.0 embodied_emissions = 0.0 model_info = "N/A" ### Conversion factors kg_per_mile = 0.348 embodied_conversion_factor = 0.0289 if "is_shared" not in st.session_state: st.session_state["is_shared"] = False electricity = pd.read_csv(electricity_url) servers = pd.read_csv(server_url) embodied_gpu = pd.read_csv(embodied_gpu_url) #st.image('images/MIT_carbon_image_narrow.png', use_column_width=True, caption = 'Image credit: ') st.title("AI Carbon Calculator") st.markdown('## Estimate your AI model\'s CO2 carbon footprint! 🌎🖥️🌎') st.markdown('### Calculating the carbon footprint of AI models can be hard... this tool is here to help!') st.markdown('##### Use the calculators below to calculate different aspects of your model\'s carbon footprint' \ 'and don\'t forget to share your data to help the community better understand the carbon emissions of AI!') st.markdown('### Dynamic Emissions 🚀') st.markdown('##### These are the emissions produced by generating the electricity necessary for powering model training.') with st.expander("Calculate the dynamic emissions of your model"): col1, col2, col3, col4, col5 = st.columns(5) with col1: hardware = st.selectbox('Hardware used', TDP['name'].tolist()) gpu_tdp = TDP['tdp_watts'][TDP['name'] == hardware].tolist()[0] st.markdown("Different hardware has different efficiencies, which impacts how much energy you use.") with col2: num_gpus = st.text_input('Number of GPUs/CPUs/TPUs used', value = 16) st.markdown('If you can\'t find your hardware in the list, select the closest similar model.') with col3: training_time = st.number_input('Total training time (in hours)', value = 0.0) st.markdown('You can find this number in your training logs or TensorBoards') with col4: provider = st.selectbox('Provider used', providers) st.markdown('If you can\'t find your provider here, select "Local/Private Infrastructure".') with col5: if provider != 'Local/Private Infastructure': provider_instances = instances['region'][instances['provider'] == provider.lower()].unique().tolist() region = st.selectbox('Region used', provider_instances) carbon_intensity = instances['impact'][(instances['provider'] == provider.lower()) & (instances['region'] == region)].tolist()[0] else: carbon_intensity = st.number_input('Carbon intensity of your energy grid, in grams of CO2 per kWh') st.markdown('You can consult a resource like the [IEA](https://www.iea.org/countries) or ' ' [Electricity Map](https://app.electricitymaps.com/) to get this information.') dynamic_emissions = round(gpu_tdp * float(num_gpus)*training_time * carbon_intensity/1000000) st.metric(label="Dynamic emissions", value=str(dynamic_emissions)+' kilograms of CO2eq') st.info('This is roughly equivalent to '+ str(round(dynamic_emissions/kg_per_mile,1)) + ' miles driven in an average US car' ' produced in 2021. [(Source: energy.gov)](https://www.energy.gov/eere/vehicles/articles/fotw-1223-january-31-2022-average-carbon-dioxide-emissions-2021-model-year)') st.markdown('### Experimental Emissions 👩‍🔬') st.markdown('##### These are the emissions produced by generating the electricity necessary for powering the experiments and tests needed to pick your final model architecture ' 'and parameters.') with st.expander("Calculate the experimental emissions of your model"): #st.info('Consult your training logs to figure out how many ablations, baselines and experiments were run before converging on the final model.') experimentation_time = st.number_input(label='Number of hours of experimentation (including ablations, baselines and evaluation)', value=training_time) st.info('As a baseline, language models such as [OPT](https://arxiv.org/pdf/2205.01068.pdf) and [BLOOM](https://arxiv.org/abs/2211.02001)' ' found that experimentation roughly doubles the amount of compute used by training the model itself.') experimental_emissions = round(gpu_tdp * (experimentation_time) * carbon_intensity/1000000) st.metric(label="Experimental emissions", value=str(experimental_emissions)+' kilograms of CO2eq') st.markdown('### Datacenter (Overhead) Emissions 🌐') st.markdown('##### These are the emissions produced by generating the electricity needed to power the rest of the infrastructure' 'used for model training -- the datacenter, network, heating/cooling, storage, etc.') with st.expander("Calculate the datacenter emissions of your model"): st.info('A proxy often used to reflect idle emissions is PUE (Power Usage Effectiveness), which represents ' ' the ratio of energy used for computing overheads like cooling, which varies depending on the data center.') pue = instances['PUE'][(instances['provider'] == provider.lower()) & (instances['region'] == region)].tolist()[0] source = instances['PUE source'][(instances['provider'] == provider.lower()) & (instances['region'] == region)].tolist()[0] if math.isnan(pue) == True: if provider != 'Local/Private Infastructure': st.markdown('##### The exact information isn\'t available for this datacenter! We will use your provider\'s average instead, which is:') if provider == 'AWS': pue = 1.135 st.markdown('#### ' + str(pue)+ " [(source)](https://www.cloudcarbonfootprint.org/docs/methodology/)") elif provider == 'GCP': pue = 1.1 st.markdown('#### ' + str(pue) + " [(source)](https://www.google.ca/about/datacenters/efficiency/)") elif provider == 'AZURE': pue = 1.185 st.markdown('#### ' + str(pue) + " [(source)](https://www.cloudcarbonfootprint.org/docs/methodology/)") elif provider == 'OVH': pue = 1.28 st.markdown('#### ' + str(pue) + " [(source)](https://corporate.ovhcloud.com/en-ca/sustainability/environment/)") elif provider == 'SCALEWAY': pue = 1.35 st.markdown('#### ' +str(pue) + " [(source)](https://pue.dc3.scaleway.com/en/)") else: st.markdown('##### Try to find the PUE of your local infrastructure. Otherwise, you can use the industry average, 1.58:') pue = st.slider('Total number of GPU hours', value = 1.58) else: st.markdown('##### The PUE of the datacenter you used is: '+ str(pue) + ' [(source)]('+source+')') pue_emissions = round((experimental_emissions+ dynamic_emissions)*pue) st.metric(label="Dynamic and experimental emissions, considering PUE", value=str(pue_emissions)+' kilograms of CO2eq') st.markdown('### Embodied Emissions 🖥️🔨') st.markdown('##### These are the emissions associated with the materials and processes involved in producing' ' the computing equipment needed for AI models.') with st.expander("Calculate the embodied emissions of your model"): st.markdown('These are the trickiest emissions to track down since a lot of the information needed is missing.') st.markdown('##### Based on the number of GPUs and training time you indicated above, we can estimate that your model\'s embodied emissions are approximately: ') hardware_type = TDP['type'][TDP['name'] == hardware].tolist()[0] if hardware_type == 'cpu': embodied_type = embodied_gpu['Value'][embodied_gpu['Ratio']=='Manufacturing emissions per additional CPU (kgCO₂eq)'].tolist()[0] elif hardware_type == 'gpu' or hardware_type == 'tpu': embodied_type = embodied_gpu['Value'][embodied_gpu['Ratio']=='Manufacturing emissions per additionnal GPU Card (kgCO₂eq)'].tolist()[0] embodied_emissions = round(int(embodied_type)*embodied_conversion_factor*float(num_gpus)*training_time/1000,1) st.metric(label="Embodied emissions", value=str(embodied_emissions)+' kilograms of CO2eq') st.markdown('This is a high-level estimate based on an hourly manufacturing emissions conversion factor (linearly ammortised) of 0.0289 [(source)](https://docs.google.com/spreadsheets/d/1DqYgQnEDLQVQm5acMAhLgHLD8xXCG9BIrk-_Nv6jF3k/).') st.markdown('### Model Information ℹ️') st.markdown('##### If you want to share the link to your model code or paper, please do so below! Otherwise, your submission will be anonymous.') model_info = st.text_input(label= "Enter a link to your model (optional)") m = st.markdown(""" """, unsafe_allow_html=True) buttoncol1, buttoncol2, buttoncol3 = st.columns(3) with buttoncol2: if not st.session_state["is_shared"]: submitted = st.button(label="Share my CO2 data!", on_click = lambda *args: write_to_csv(hardware, gpu_tdp, num_gpus, training_time, provider, carbon_intensity, dynamic_emissions, experimentation_time, experimental_emissions, pue, pue_emissions, embodied_type, embodied_emissions, model_info)) else: st.info('Thank you! Your data has been shared in https://huggingface.co/datasets/sasha/co2_submissions.') st.markdown('### Methodology') with st.expander("More information about our Methodology"): st.markdown('Building on the work of the [ML CO2 Calculator](https://mlco2.github.io/impact/), this tool allows you to consider' ' other aspects of your model\'s carbon footprint based on the LCA methodology.') st.markdown('We considered all of these aspects when calculating the CO2 emissions of BLOOM 🌸, a 176-billion parameter language model [(see our preprint!)](https://arxiv.org/abs/2211.02001)'')') st.image('images/LCA_CO2.png', caption='The LCA methodology - the parts in green are those we focus on.')