AI, according to Mistral AI's environmental report, reveals a significant demand for resources, likened to a ravenous, insatiable creature in terms of energy and water consumption.
In a groundbreaking move, Mistral AI has published a peer-reviewed report detailing the environmental impact of its Mistral Large 2 LLM, a generative AI model. The report, which emphasizes transparency and comparability, advocates for standardized, internationally recognized frameworks for reporting and measuring AI's environmental footprint.
The key findings of the report are centred around three core impact metrics:
- Environmental impact of training the model: The report quantifies the energy consumption, water consumption, and carbon emissions associated with training the model. For instance, training the 123 billion parameter model produced approximately 20 kilotons of CO2 equivalents (CO2e) and consumed 281,000 cubic meters of water.
- Ongoing environmental costs of running/inference of the model: The report also highlights the environmental costs associated with running the model, accounting for 85.5% of GHG emissions and 91% of water consumption. During inference, the model consumed about 45 ml of water and generated about 1.14 grams of CO2e.
- Portion of the model’s lifespan spent on training versus inference: This metric provides context for the amortization of resource use over the model's lifespan.
Mistral AI suggests that users, developers, and policy makers find these first two details essential, while the third could be an internal metric or released to the public.
The report also underscores the importance of life-cycle assessment (LCA) approaches, covering the full span of the model’s life—including hardware manufacturing, training, and inference. Such assessments show that inference can contribute 65-90% of total lifecycle emissions due to massive deployment scale.
To address water consumption, Mistral argues that cooling towers employed by AI datacenters can be problematic in drought-prone regions. Building training models in cool climates with renewable energy can help reduce carbon footprint and water consumption.
The report further advocates for the use of granular, time- and location-specific emission factors for data center energy usage, incorporating probabilistic modeling and sensitivity analysis to pinpoint major emission drivers.
In addition, Mistral calls for public publication of environmental impact reports by AI developers to foster industry-wide transparency and enable scoring systems for carbon-, water-, and material-intensity of AI models. This, in turn, would help users and policymakers choose sustainable options.
The report's findings closely align with prior research into AI's water consumption habits. Techniques like speculative decoding or sparse model architectures could further reduce AI's environmental impact.
As the field of AI continues to evolve, the need for standardized, internationally recognized frameworks for reporting and measuring environmental impact becomes increasingly important. This area remains in early development with ongoing calls for broader adoption and refinement of these standards.
- In the report, Mistral AI highlights the significance of life-cycle assessment (LCA) approaches, including the environmental impact of training the Mistral Large 2 LLM, ongoing environmental costs of running the model, and the portion of its lifespan spent on training versus inference.
- The report suggests that AI developers should publish environmental impact reports publicly to foster industry-wide transparency and enable scoring systems for the carbon-, water-, and material-intensity of AI models.
- The findings of the Mistral AI report align with prior research into AI's water consumption habits, emphasizing the need for granular, time- and location-specific emission factors for data center energy usage.
- In response to climate-change concerns and the growing importance of environmental-science, the report advocates for the incorporation of probabilistic modeling and sensitivity analysis to pinpoint major emission drivers, using techniques like speculative decoding or sparse model architectures to further reduce AI's environmental impact.
