Adding carbon footprint information

Calculated the carbon footprint of different stages of IDEFICS training + experimentation based on information provided by

@VictorSanh

and

@stas

!

README.md

@@ -317,6 +317,32 @@ The IDEFICS models were trained on an AWS SageMaker cluster with 8x80GB A100 GPU
 The training software is built on top of HuggingFace Transformers + Accelerate, and [DeepSpeed ZeRO-3](https://github.com/microsoft/DeepSpeed) for training, and [WebDataset](https://github.com/webdataset/webdataset) for data loading.
 
 
+## Environmental Impact
+
+We distinguish the 3 phases of the creation of IDEFICS and report our carbon emissions separately for each one of them:
+
+*Preliminary experimentation*
+- **Hardware Type:** Intel Cascade Lake CPUs, NVIDIA V100 and A100 GPUs
+- **Hours used:** 460,000 CPU hours, 385,000 V100 GPU hours, and 300,000 A100 GPU hours
+- **Cloud Provider:** N/A (Jean Zay cluster)
+- **Compute Region:** France (57g CO2eq/kWh)
+- **Carbon Emitted:** 16,714 kgs of CO2eq
+
+*IDEFICS-80B pretraining* 
+- **Hardware Type:** 512 NVIDIA A100 GPUs
+- **Hours used:** 672 hours (28 days)
+- **Cloud Provider:** AWS
+- **Compute Region:** US-West 2 (288g CO2eq/kWh)
+- **Carbon Emitted:** 39,498 kgs of CO2eq
+
+*IDEFICS-80b-instruct finetuning*
+- **Hardware Type:** 384 NVIDIA A100 GPUs
+- **Hours used:** 72 hours (3 days)
+- **Cloud Provider:** AWS
+- **Compute Region:** US-West 2 (288g CO2eq/kWh)
+- **Carbon Emitted:** 3,174 kgs of CO2eq
+
+
 # Bias, Risks, and Limitations
 
 Significant research has explored bias and fairness issues with language models (see, e.g., [Sheng et al. (2021)](https://aclanthology.org/2021.acl-long.330.pdf) and [Bender et al. (2021)](https://dl.acm.org/doi/pdf/10.1145/3442188.3445922)).